UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

System for vessel characterization : development and evaluation with application to deep vein thrombosis… Guerrero, Julian 2008

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2008_fall_guerrero_julian.pdf [ 27.99MB ]
Metadata
JSON: 24-1.0065501.json
JSON-LD: 24-1.0065501-ld.json
RDF/XML (Pretty): 24-1.0065501-rdf.xml
RDF/JSON: 24-1.0065501-rdf.json
Turtle: 24-1.0065501-turtle.txt
N-Triples: 24-1.0065501-rdf-ntriples.txt
Original Record: 24-1.0065501-source.json
Full Text
24-1.0065501-fulltext.txt
Citation
24-1.0065501.ris

Full Text

System for Vessel Characterization: Development and Evaluation with Application to Deep Vein Thrombosis Diagnosis by Julian Guerrero  B.Eng., Instituto Tecnol´ogico y de Estudios Superiores de Monterrey, Campus Morelos, 1999 M.A.Sc., The University of British Columbia, 2003  A THESIS SUBMITTED IN PARTIAL FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF Doctor of Philosophy in The Faculty of Graduate Studies (Electrical and Computer Engineering)  The University Of British Columbia (Vancouver) August, 2008 c Julian Guerrero, 2008  Abstract A system for vessel characterization aimed at detecting deep vein thrombosis (DVT) in the lower limbs has been developed and evaluated using ultrasound image processing, location and force sensors measurements, blood flow information and a protocol based on the current clinical standard, compression ultrasound. The goal is to provide an objective and repeatable system to measure DVT in a rapid and standardized manner, as this has been suggested in the literature as an approach to improve overall detection of the disease. The system uses a spatial Kalman filter-based algorithm with an elliptical model in the measurement equation to detect vessel contours in transverse ultrasound images and estimate ellipse parameters, and temporal constant velocity Kalman filters for tracking vessel location in real-time. The vessel characterization also comprises building a 3–D vessel model and performing compression and blood flow assessments to calculate measures that indicate the possibility of DVT in a vessel. A user interface designed for assessing a vessel for DVT was also developed. The system and components were implemented and tested in simulations, laboratory settings, and clinical settings. Contour detection results are good, with mean and rms errors ranging from 1.47–3.64 and 3.69–9.67 pixels, respectively, in simulated and patient images, and parameter estimation errors of ∼5%. Experiments showed errors of 3-5 pixels for the tracking approaches. The measures for DVT were evaluated, independently and integrated in the system. The complete system was evaluated, with sensitivity of 67–100% and specificity of 50–89.5%. System learnability and memorability were evaluated in a separate user study, with good results. Contributions include a segmentation approach using a full parameter ellipse model in an extended Kalman filter, incorporating multiple measurements, an alternate sampling method for faster parameter convergence and application–specific initialization, and a tracking approach that includes a sub–sampled sum of absolutes similarity calculation and a method to detect vessel bifurcations using flow data. Further contributions include an integrated system for DVT detection that can combine ultrasound B-mode, colour flow and elastography images for vessel characterization, a system interface design focusing on usability that was evaluated with medical professionals, and system evaluations through multiple patient studies.  ii  Table of Contents Abstract  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  ii  Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  iii  List of Tables  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  vii  List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  ix  List of Symbols  xv  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  List of Abbreviations Glossary  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  xvi  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii  Acknowledgments Dedication  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  xix  1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1  1.1  Motivation  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  2  1.2  Research Goals and Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  3  1.2.1  Functional Specification of a DVT Screening System . . . . . . . . . . . . .  5  Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  6  1.3  2 Literature Review  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  7  2.1  Deep Venous Thrombosis (DVT) . . . . . . . . . . . . . . . . . . . . . . . . . . . .  7  2.2  DVT Detection Methods  2.3  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  10  2.2.1  Contrast Venography (CV) . . . . . . . . . . . . . . . . . . . . . . . . . . .  11  2.2.2  Ultrasound Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  12  2.2.3  DVT Screening System . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  20  State of the Art  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  23  2.3.1  Image Processing  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  23  2.3.2  Location and Force Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . .  26  2.3.3  3-D Vessel Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  27  2.3.4  Characterization of Physical Properties . . . . . . . . . . . . . . . . . . . .  27  2.4  Usability  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  28  2.5  Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  30 iii  3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1  3.2  3.3  3.4  3.5  31  Contour Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  31  3.1.1  Common Model Features . . . . . . . . . . . . . . . . . . . . . . . . . . . .  32  3.1.2  3–Parameter Ellipse Models  . . . . . . . . . . . . . . . . . . . . . . . . . .  32  3.1.3  5–Parameter Ellipse Models  . . . . . . . . . . . . . . . . . . . . . . . . . .  35  3.1.4  Staggered Radial Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . .  38  3.1.5  Angle-Dependant Measurement Covariance . . . . . . . . . . . . . . . . . .  39  3.1.6  Models with Multiple Measurements . . . . . . . . . . . . . . . . . . . . . .  39  3.1.7  Control Parameters  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  44  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  46  3.2.1  Delayed Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  46  3.2.2  Modifications to the Measurement Equation  . . . . . . . . . . . . . . . . .  48  3.2.3  Interacting Multiple Models  . . . . . . . . . . . . . . . . . . . . . . . . . .  53  3.2.4  2-D vs. 3-D Coordinates in Tracking Model  3.2.5  Seed Point Persistence  3.2.6  Seed Correction Using Sum of Absolutes  3.2.7  Tracking Vessel Bifurcations  3.2.8  Vessel Specific Tracking for DVT  3.2.9  Control Parameters  Vessel Tracking  . . . . . . . . . . . . . . . . .  54  . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  54  . . . . . . . . . . . . . . . . . . .  55  . . . . . . . . . . . . . . . . . . . . . . . . . .  57  . . . . . . . . . . . . . . . . . . . . . . .  59  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  60  Model Construction and Display . . . . . . . . . . . . . . . . . . . . . . . . . . . .  60  3.3.1  Contour Averaging vs. Contour Insertion . . . . . . . . . . . . . . . . . . .  61  3.3.2  Dynamic Contour Sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . .  62  3.3.3  Mapping Multiple Data to Model Surface . . . . . . . . . . . . . . . . . . .  62  3.3.4  Branching Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  63  3.3.5  Additional Virtual Objects . . . . . . . . . . . . . . . . . . . . . . . . . . .  66  3.3.6  Control Parameters  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  68  Vessel Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  69  3.4.1  Compression Assessment  . . . . . . . . . . . . . . . . . . . . . . . . . . . .  69  3.4.2  Use of Blood Flow Data  . . . . . . . . . . . . . . . . . . . . . . . . . . . .  71  3.4.3  Local Contour Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . .  72  3.4.4  Control Parameters  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  73  System Integration and User Interface . . . . . . . . . . . . . . . . . . . . . . . . .  73  3.5.1  Ultrasonix Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  74  3.5.2  Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  76  3.5.3  User Interface  76  3.5.4  Scanning Overview  3.5.5  Proposed Examination Protocol  3.5.6  Vessel Viewer  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  80  . . . . . . . . . . . . . . . . . . . . . . . .  81  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  83  iv  4 Experiments and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1  4.2  Contour Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  84  4.1.1  3–Parameter Ellipse Model . . . . . . . . . . . . . . . . . . . . . . . . . . .  84  4.1.2  5–Parameter Ellipse Model . . . . . . . . . . . . . . . . . . . . . . . . . . .  95  Vessel Tracking 4.2.1  4.3  4.4  4.5  84  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104  Tracking Vessel Bifurcations  DVT Screening System  . . . . . . . . . . . . . . . . . . . . . . . . . . 106  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110  4.3.1  ROC Curves for TAR and Slope Criteria  . . . . . . . . . . . . . . . . . . . 110  4.3.2  System Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111  4.3.3  Clinical Pilot Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111  Usability Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 4.4.1  Usability Evaluation Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . 115  4.4.2  Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118  Clinical Study 4.5.1  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127  Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129  5 Discussion and Conclusions  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133  5.1  Contour Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133  5.2  Vessel Tracking  5.3  DVT Screening System  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138  5.3.1  Usability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140  5.3.2  Clinical Evaluations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142  5.3.3  Costs of a DVT Screening System . . . . . . . . . . . . . . . . . . . . . . . 144  5.4  Contributions  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145  5.5  Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148  Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151  Appendices A Kalman Filters  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161  A.1 Extended Kalman Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 A.1.1 Parameter Estimation using EKFs . . . . . . . . . . . . . . . . . . . . . . . 162 A.2 Interacting Multiple Models (IMM)  . . . . . . . . . . . . . . . . . . . . . . . . . . 163  B Contour Detection and Vessel Tracking . . . . . . . . . . . . . . . . . . . . . . . . 166 B.1 Contour Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 B.1.1 Derivative Method for Contour Detection Using a 3–Parameter Ellipse Model 166 B.1.2 Ellipsoid Model for Parameter Estimation and Boundary Detection B.2 Vessel Tracking Models  . . . . 167  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171  B.2.1 Velocity in the System B.3 Tracking Vessel Bifurcations  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 v  B.3.1 Grayscale or B-mode Single Image Measures . . . . . . . . . . . . . . . . . 172 B.3.2 Grayscale or B-mode Multiple Image Measures . . . . . . . . . . . . . . . . 174 B.3.3 Bifurcation Detection with Flow Image Measures B.3.4 Experiments and Validation  . . . . . . . . . . . . . . 174  . . . . . . . . . . . . . . . . . . . . . . . . . . 174  C Cost Estimate and Implications of a DVT Screening System  . . . . . . . . . . 183  C.1 Estimate of the Cost of a Prototype Screening System . . . . . . . . . . . . . . . . 183 C.2 Estimate of the Cost of a Commercial Screening System . . . . . . . . . . . . . . . 184 C.3 Cost Implications of a DVT Screening System  . . . . . . . . . . . . . . . . . . . . 187  C.3.1 Screening Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 C.3.2 Hospital Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 C.3.3 Cost Comparison  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188  C.3.4 Cost of False Negatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 D Usability and System Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196 D.1 Scan Protocol - DVT Screening System . . . . . . . . . . . . . . . . . . . . . . . . 197 D.2 User Instruction Sheet - Learnability  . . . . . . . . . . . . . . . . . . . . . . . . . 198  D.3 User Instruction Sheet - Memorability . . . . . . . . . . . . . . . . . . . . . . . . . 204 D.4 Questionnaire - User Satisfaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 D.5 Consent Form, Usability Study - User Participant  . . . . . . . . . . . . . . . . . . 208  D.6 Consent Form, Usability Study - Volunter Participant . . . . . . . . . . . . . . . . 212 D.7 Consent Form, Patient Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 E Ethics Approval Forms  . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220  E.1 UBC CREB approval (2003) of pilot study of DVT screening system  . . . . . . . 221  E.2 UBC CREB renewal (2004) of pilot study of DVT screening system . . . . . . . . 222 E.3 UBC CREB renewal (2005) of pilot study of DVT screening system . . . . . . . . 223 E.4 VCHA approval (2003) of pilot study of DVT screening system . . . . . . . . . . . 224 E.5 VCHA approval (2004) of pilot study of DVT screening system . . . . . . . . . . . 225 E.6 UBC CREB approval (2007) of study of DVT screening system  . . . . . . . . . . 226  E.7 UBC CREB renewal with amendments (2007) of study of DVT screening system . 228 E.8 UBC CREB approval of amendments (2008) of study of DVT screening system . . 230 E.9 VCHA approval (2008) of study of DVT screening system . . . . . . . . . . . . . . 232  vi  List of Tables 4.1  Convergence benchmarks for 3–parameter ellipse models. Shows the average number of iterations required to reach each benchmark. . . . . . . . . . . . . . . . . . . . .  85  4.2  Overall parameter estimation results - simulated images. . . . . . . . . . . . . . . .  89  4.3  Detected ellipse parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  90  4.4  Contour detection results - simulated images, errors per expert, using validation √ threshold = 1.5 · rmax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  91  4.5  Contour detection results - simulated images, errors per expert, using validation threshold = 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  91  4.6  Parameter estimation and contour detection results for images with attenuation. .  91  4.7  Parameter estimation and contour detection results for images with varying signalto-noise ratios (SNRs). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  92  4.8  Overall results - patient images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  93  4.9  Overall results, in pixels - patient images. . . . . . . . . . . . . . . . . . . . . . . .  94  4.10 Contour detection results - patient images, errors per expert, using validation thresh√ old = 1.5 · rmax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  95  4.11 Contour detection results - Patient images, errors per expert, using validation threshold = 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  95  4.12 Convergence benchmarks for 5–parameter ellipse models. Shows the average number of iterations required to reach each benchmark. . . . . . . . . . . . . . . . . . . . .  96  4.13 Standard deviation of errors for 5–parameter ellipse models. . . . . . . . . . . . . .  96  4.14 Errors between estimated ellipse parameters and true values. . . . . . . . . . . . . 100 4.15 Algorithm execution times (ms). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4.16 Errors between detected contour and expert tracings for simulated images. . . . . . 103 4.17 Errors between detected contour and expert tracings for patient images. . . . . . . 103 4.18 Errors and execution time for different SSA sizes. . . . . . . . . . . . . . . . . . . . 105 4.19 Detecting vessel bifurcations using flow data, first data set . . . . . . . . . . . . . . 109 4.20 Detecting vessel bifurcations using flow data, second data set . . . . . . . . . . . . 109 4.21 Execution time (ms) - DVT screening system. . . . . . . . . . . . . . . . . . . . . . 111 4.22 Minimum applied force during compression exams. . . . . . . . . . . . . . . . . . . 114 4.23 Maximum applied force during compression exams. . . . . . . . . . . . . . . . . . . 114 4.24 Compression assessment results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 4.25 Learnability Indices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 4.26 Learnability Indices - All Users. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 4.27 Questionnaire Statements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 vii  4.28 Overall Questionnaire Responses. See Table 4.27 for key to question numbers. . . . 124 4.29 Questionnaire Responses - User Comments. . . . . . . . . . . . . . . . . . . . . . . 126 4.30 User Errors, Round 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.31 User Errors, Round 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128 4.32 User Errors - ‘Talking Out Loud’ Comments. . . . . . . . . . . . . . . . . . . . . . 129 4.33 System Errors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.34 J values for vessel measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.35 Mean (µ), standard deviation (std. dev.), maximum and minimum for optimal feature, for positive and negative DVT classes. . . . . . . . . . . . . . . . . . . . . 132 B.1 Convergence benchmarks for the ellipsoid model. Shows the average number of iterations required to reach each benchmark. . . . . . . . . . . . . . . . . . . . . . 171 B.2 Single image measures calculated for vessel bifurcation classification. . . . . . . . . 175 B.3 Multiple image measures calculated for vessel bifurcation classification. . . . . . . . 178 C.1 Cost estimates of the prototype DVT screening system . . . . . . . . . . . . . . . . 184 C.2 Cost estimates of a commercial DVT screening system . . . . . . . . . . . . . . . . 185 C.3 Screening costs per patient. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 C.4 Costs comparison per 1000 patients. . . . . . . . . . . . . . . . . . . . . . . . . . . 189 C.5 Estimate of time savings per 1000 patients. . . . . . . . . . . . . . . . . . . . . . . 189 C.6 Overall cost difference per 1000 patients. . . . . . . . . . . . . . . . . . . . . . . . . 189 C.7 Cost estimate of detecting and treating DVT per patient. . . . . . . . . . . . . . . 193 C.8 Cost estimate of detecting and treating PE per patient. . . . . . . . . . . . . . . . 193 C.9 Cost estimate of patient treatment based on DVT detection sensitivity.  . . . . . . 195  viii  List of Figures 2.1  Venous valves shown in a closed position when pockets are formed behind the cusps, with normal venous flow (blue) and blocked return flow (green). . . . . . . . . . . .  8  2.2  Diagram showing the major vessels of the deep venous system. . . . . . . . . . . .  9  2.3  Image sequence showing the compressibility of a SFV (bottom) as force is applied (left to right) with an ultrasound probe.The artery remains with a similar area throughout (top). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  2.4  12  Sensorized ultrasound probe used in the DVT screening system. Schematic (a) shows the inner and outer shells, and force /torque sensor, while (b) shows a user grasping the outer shell. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  2.5  Sensorized ultrasound probe with exposed inner shell, force–torque sensor casing, and location sensor mounted on a Plexiglas rod.  2.6  21  . . . . . . . . . . . . . . . . . . .  22  Example of an IVUS image of a vessel (a) and a vessel imaged using a linear US array (b). Key differences include the coordinate system, resolution, and image artifacts (a small black circle at the centre of (a) where the catheter is). . . . . . .  3.1  A contour in an ultrasound image can be approximated by using an elliptical model with parameters a, b and φ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  3.2  33  Oscillating estimates of a (‘+’) and b (‘– –’), and true parameter value (‘–’, 70 for both a and b) over 4 cycles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  3.3  25  35  Ellipse with parameters a, b, φ, xe and ye . Distance from seed point (xc , yc ) to contour described by ρ and angle α (a) or θ (b). In (a), if ellipse is centered at (xc , yc ), then r = ρ, and θ = α. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  3.4  Image data displayed in radial coordinates using a small angular step size (b) and a large step size (c). Detected contour dipslayed with ‘o’, and M = 8, N = 64. . .  3.5  36  Measurement covariance vs.  angle.  38  Minimum covariance values occur at θ =  π/2, 3π/2 corresponding to top and bottom of vessel, while maximum values occur at θ = π, 2π corresponding to the vessel sides. . . . . . . . . . . . . . . . . . . . 3.6  40  Construction of mask for similarity measure. Given the contour location rki−1 from the previous frame i for the current radius k, a mask is constructed from the pixel values of the previous image using a predetermined width 2Ws + 1. In this example, W = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  42  ix  3.7  Examples of strain images (a), (b), (c) with B-mode data (d), (e), (f) acquired at the same time. A hard inclusion can clearly be seen in the strain image (a) although it does not show up as clearly on the B-mode (d). Also shown are images of the carotid artery. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  3.8  43  Example of vessels in an ultrasound image. Notice how the top and bottom boundaries of the vessels are brighter and more clearly defined than the lateral boundaries. Also included is the orientation of the coordinate axes. . . . . . . . . . . . . . . . .  3.9  45  Tracking the seed point. An estimate x ˆi of the vessel centre (xci , yci ) along trajectory s is calculated for each image plane Πi using a constant velocity model. The previously obtained vessel centre ci−1 is used as measurement. . . . . . . . . . . . .  47  3.10 Projection of the estimated seed point x ˆi onto the current image plane Πi . . . . .  48  3.11 The location with minimum brightness in an image region is found using the sum of absolutes and all (a), half (b), or one fifth (c) of the pixels in the region. . . . .  55  3.12 Including weights in the SSA calculation. Examples of image data without weights (a), linearly increasing weights from the centre (b) and quadraticly increasing weights from the centre (c) are shown, where weight values shown in (d)-(f) taken from the diagonal in (a)-(c). In this example, weight values are in the range of 1 to 16 (image brightness is in the range 0 to 255) and image is displayed in false colour. Notice the increase in image intensities in the weighted images, especially near the edges.  56  3.13 Minimum brightness value using the SSA approach on an unweighted image (a) and a weighted images (b), (c). Notice how the locations of minimum brightness (dark blue) in the unweighted data are shifted toward the edges while the areas of minimum brightness are much more localized in the weighted data, and closer to the desired location (centre of the image). . . . . . . . . . . . . . . . . . . . . . . .  57  3.14 A vessel bifurcation in B-mode with the detected contour (white circles) from one branch (a), with corresponding colour flow data (b). The thresholded and labeled data clearly shows two regions, one inside the contour and the other outside. The second flow region is used to initialize a new contour. . . . . . . . . . . . . . . . . .  58  3.15 A seed point for artery (A) is tracked and used to calculate an estimated seed point for a vein (B), which is placed outside the previous contour of the artery for a final vein seed point (C). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  59  3.16 A polygonal mesh is generated to describe the surface of the 3-D vessel model (a), and data describing vessel compressibility (or other characteristics) can be mapped to the vessel surface (b). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  61  3.17 Model Mesh. 3-D vessel model is constructed by creating polygons using detected contour points as vertices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  62  3.18 A new contour is inserted into a 3-D vessel model if it is far enough from all previous contours (a),(c), but is averaged with an existing contour if it is close enough to that contour (b),(d). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  63  x  3.19 Examples of data mapped as colour to the surface of a 3-D vessel model. A single colour for each contour mapped to the vessel surface (a), and multiple colours for each contour using different colourmaps (b), (c). . . . . . . . . . . . . . . . . . . .  64  3.20 Constructing a branching 3-D vessel model. A 3-D model is generated, and the location of the base contour for the new branch is selected (a),(b). The model building procedure is then continued, with the added contours forming a new vessel branch (c),(d). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  65  3.21 Constructing a 3-D vessel model with unconnected branches. Independent seed tracking and contour instances were used to generate two unconnected vessel branches. 66 3.22 Additional 3-D models on the 3-D display. (a) Vessel model with mapped compression data and text labels, reset point and colour bar. (b) Vessel model and ultrasound plane shown with skin surface model and reset point. (c) Vessel model with ‘vessel origin’ sphere. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  67  3.23 Hardware diagram of the DVT screening system. . . . . . . . . . . . . . . . . . . .  74  3.24 Ultrasonix Sonix  R  PC-based ultrasound with DVT screening system. . . . . . . .  75  3.25 The screening system interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  77  3.26 The ‘Help with Scan’ dialog. A short version of scanning instructions for use with the DVT screening system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  78  3.27 The ‘Help’ dialog. A more detailed version of scanning instructions for use with the DVT screening system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  79  3.28 The ‘Flow Help’ dialog. Detailed instructions for flow assessment using the DVT screening system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  80  3.29 Vessel Viewer Software. Clockwise from left, navigation display, 3–D vessel model display, compression criteria display, and ultrasound image with detected contour. 4.1  Final ellipse parameter (a), (b) and contour (c), (d) errors obtained during convergence tests for the 3–parameter ellipse model using the output approach. . . . . . .  4.2  82  86  Examples of contour and parameter estimation. Simulated data used to test contour detection, with expert tracing (a) and detected contour (b) with detected contours (white circles). The estimated semi-major axis a and semi-minor axis b are presented as a solid and dashed line, respectively, oriented according to the estimated φ. . . .  4.3  Location of seed points. Example of seed points (.) generated for a simulated image. Expert tracing shown in gray. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  4.4  87 88  Patient data (images cropped to 200×200 pixels) with examples of detected contours (white circles) and estimated ellipses. The estimated semi-major axis a and semiminor axis b are presented as a solid and dashed line, respectively, oriented according to the estimated φ. Expert tracings presented in green. (a), (b), (d) and (f) reprinted with permission [123], (c) and (e) obtained from laboratory tests. Even though expert tracings are not elliptical, the detected contours correspond to the vessel shapes illustrating that the contour does not have to be elliptical even though an ellipse model is used. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  94 xi  4.5  Final ellipse parameter errors obtained during convergence tests for the 5–parameter ellipse model using the output approach. . . . . . . . . . . . . . . . . . . . . . . . .  4.6  Final contour errors obtained during convergence tests for the 5–parameter ellipse model using the output approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . .  4.7  98  Final contour errors obtained during convergence tests for the 5–parameter ellipse model using the output approach and large step size. Full error range not shown. .  4.9  97  Final ellipse parameter errors obtained during convergence tests for the 5–parameter ellipse model using the output approach and large step size. . . . . . . . . . . . . .  4.8  97  98  Example of parameter convergence using a large angular step and 64 radii. Values of estimated contour and parameters (blue) are shown along with true values (black). Notice the large contour errors at several locations, which correspond to the radii that were first estimated (contour radii have been re-ordered). . . . . . . . . . . . .  99  4.10 Detection results on simulated (a) and patient (b) data with contour ‘+’ and ellipse ’– –’ using a 3–parameter ellipse algorithm, and contour ‘o’ and ellipse ’–’ using a 5–parameter ellipse algorithm, both with the same initial seed point. The results of the 3–parameter algorithm are clearly not adequate, while the results from the 5–parameter version are quite good. . . . . . . . . . . . . . . . . . . . . . . . . . . 101 4.11 Mean contour errors (pix) in simulated (a) and patient (b) images, and ellipse area errors (%) in simulated (c) and patient (d) images, for 3 (3PAR) and 5 (5PAR) parameter ellipse models, and a 5 parameter model with large step size (5PAR-l). . 102 4.12 Final ellipsoid parameter errors obtained during convergence tests for the ellipsoid model using the output approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.13 Seed tracking results. A seed point was tracked in a slow step (a), and a medium and fast step (Figure 4.14). Manual segmentation is represented by the solid line, the previous tracking method by (+) and the method presented herein by (.), and the x (above) and y (below) coordinates are presented. Errors (b) are also shown.  106  4.14 Seed tracking results. A seed point was tracked in a slow (Figure 4.13), medium (a) and fast step (b). Manual segmentation is represented by the solid line, the previous tracking method by (+) and the method presented herein by (.), and the x (above) and y (below) coordinates are presented. Errors for the medium (c) and fast (d) steps are also shown. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 4.15 SSA validation. A seed point was tracked in a slow (a) and medium step (d), using all (+), half (.), and one fifth the pixels (x). Execution times per image for the slow (b) and medium step (d) are presented, as well as respective errors (c), (f). . . . . 108 4.16 Typical human compression data consisting of normalized transverse area vs. normalized force is presented. Vein (a) transverse area almost disappears as force is applied, while artery (b) transverse area does not change over range of applied force. 110 4.17 ROC curves for TAR (a) and slope (b) venous compressibility criterion. . . . . . . 111 4.18 DVT system interface. The virtual vessel model is presented on the left, while ultrasound data with detected contour is displayed on the right. . . . . . . . . . . . 112  xii  4.19 Box plots of learnability indices L1, L2, L3 and L4, calculated from user data in rounds 1 and 2. The boxes have lines at the lower quartile, median, and upper quartile values. The whiskers are lines extending from each end of the boxes to show the extent of the rest of the data. Outliers are data with values beyond the ends of the whiskers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 4.20 Box plots of learnability indices L1, L2, L3 and L4, calculated for each group (‘NP’ = nurses, ‘US’ = US technicians, ‘MD’ = medical students) from data in rounds 1 and 2. See Figure 4.19 for box descriptions. . . . . . . . . . . . . . . . . . . . . . . 121 4.21 Box plots of scan time for conventional ultrasound compression examinations, when scanning one or two limbs. Does not include time needed for review by radiologist. See Figure 4.19 for box descriptions. . . . . . . . . . . . . . . . . . . . . . . . . . . 122 4.22 Box plots of user responses to questionnaire from rounds 1 and 2. See Table 4.27 for key to question numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 B.1 An example of tangent and perpendicular lines to the detected contour for a vessel without a bifurcation, as well as the distribution of the resulting intersection points of all perpendicular lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 B.2 A detected contour (a) for a vessel bifurcation, where only one side of the vessel was detected. In the radial image data (b), the location corresponding to the second half of the vessel appears as a gap in the image intensity (arrow). . . . . . . . . . . 173 B.3 J score for each measure (1 – 18, as detailed in Table B.2) for classifying a single B-mode image as containing a vessel bifurcation or not. . . . . . . . . . . . . . . . 177 B.4 J score for each measure (1 – 20, as detailed in Table B.3) for classifying multiple B-mode images as containing a vessel bifurcation or not. . . . . . . . . . . . . . . . 180 B.5 J score for each measure (1 – 20, as detailed in Table B.3) for classifying multiple B-mode images as containing a vessel bifurcation or not using a windowed approach. The size of the window is indicated in the graph. Notice all values are very close to 1.181 B.6 Examples of typical values of the image gap measure for an image series with and without a vessel bifurcation. The difference, where the image intensity in the top sequence tapers off, is due to ultrasound artifacts and not to the presence of vessel bifurcations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 D.1 Diagram showing the major vessels of the deep venous system. . . . . . . . . . . . 197 D.2 The screening system interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199  xiii  List of Symbols a  Semi–major axis of ellipse equation  Ak  System matrix at time k  Ak (xk )  Nonlinear system matrix at time k  b  Semi–minor axis of ellipse equation  Ck  Measurement matrix at time k  Ck (xk )  Nonlinear measurement matrix at time k  ηk  Measurement noise vector at time k  Fk+1  Jacobian of Ak (x)  Gk  Kalman gain matrix  Hk+1  Jacobian of Ck (x)  M  Number of candidate points used in radial edge detection  N  Number of equispaced radii used in contour detection  φ  Angular displacement of ellipse with respect to the image coordinate axis  Pk  State covariance at time k  Pk|k  State prediction covariance  Qk  Covariance matrix of random vector ξ k  r  Radius at angle θ from the seed point  ρi  Candidate points used in radial edge detection  Rk  Covariance matrix of random vector η k  rmax  Maximum radial search distance from the seed point xc for contour detection  Sk  Measurement prediction covariance  Tk  Variance of artificial process noise vector ς k  θ  Radius angle with respect to the image coordinate axis  Θk  Parameter vector  Uk  Covariance matrix of noise vector υ k  υk  Artificial process noise vector at time k  ςk  White, zero mean artificial process noise of ellipse parameter vector  vk  Measurement vector at time k  v ˆk  Predicted measurement  xc  x coordinate of the seed point xc for star-based contour detection algorithms  xc  Seed point for star-based contour detection algorithms  ξk  System noise vector at time k  x ˆk|k−1  Optimal prediction of xk  xk , x  State vector at time k xiv  x ˆk , x ˆk|k  Optimal filtering estimate of xk  yc  y coordinate of the seed point xc for star-based contour detection algorithms  yk  Augmented state vector  zk  Measurement residual  xv  List of Abbreviations ASM  Active shape models  AT  Anterior tibial vein  CDI  Colour Doppler imaging  CFA  Common femoral artery  CFV  Common femoral vein  CUS  Compression ultrasound  CV  Contrast venography  CW  Continuous wave (Doppler, for flow)  DOF  Degrees of freedom  DUS  Duplex ultrasonography  DVT  Deep vein thrombosis  EKF  Extended Kalman filter  IVUS  Intravascular ultrasound  LSV  Long saphenous vein  PE  Pulmonary embolism  PER  Peroneal vein  POP  Popliteal vein  PROF  Deep femoral vein or profunda femoris  PT  Posterior tibial vein  PW  Pulsed wave (Doppler, for flow)  SFA  Superficial femoral artery  SFV  Superficial femoral vein  TGC  Time gain control  US  Ultrasound  xvi  Glossary Distal  Farther from the heart  Fibrin  An insoluble protein essential to blood clotting  Hyperechoic  Region in an US image where echoes are stronger than normal or than surrounding structures  Hypoechoic  Region in an US image where echoes are weaker than normal or than surrounding structures  Left lateral decubitus  Patient lies on left side with right arm over head  Leukocyte  (White blood cell) One of the cells the body makes to help fight infections  Morbidity  A disease, condition or state. Complications directly resulting from treatment.  Negative predictive value  Ratio of true negatives over total negatives detected (true negatives plus false negatives)  Platelet  An irregular, disc-shaped element in the blood that assists in blood clotting  Positive predictive value  Ratio of true positives over total positives detected (true positives plus false positives)  Prone  Patient lies on front  Prophylaxis  A measure taken for the prevention of a disease or condition  Proximal  Nearer to the heart  Right lateral decubitus  Patient lies on right side with left arm over head  Sensitivity  Ratio of true positives over all positives (true positives plus false negatives)  Sitting  Patient is erect / semi-erect  Specificity  Ratio of true negatives over all negatives (true negatives plus false positives)  Supine  Patient lies on back  xvii  Acknowledgments First of all, I would like to thank my supervisors Tim Salcudean and James McEwen for their continuing support during this project. For giving me and this project new directions, for helping with insights, contacts, and simply working toward a useful system. I would also like to acknowledge the support of Dr. Bas Masri and Dr. Savvas Nicolaou, who assisted me greatly during this whole project, especially during the evaluations. Their help is much appreciated. My data collection could not have worked without the invaluable help of Anne Hope and Maureen Kennedy at UBC Radiology, and Jean Hong and Leah Christoff at UBC Orthopaedics. You were on the front lines and shone every day. My implementation would not have succeeded without the assistance of Neerav Patel, who in addition produced a data viewer from my specifications. Thanks for coding and helping and guiding me through the innards of the operating system. I also appreciate the help, volunteering, chats and conversations with everyone in our lab. Thank you for your contributions: Orcun for insightful discussions throughout development, Reza for his assistance with strain imaging and elastography, Sara for talking about ellipsoids, Hani for classification metrics and optimization, and everyone else for just making the RCL a great open place to work. I would also like to acknowledge the support of this project by NSERC through the IPS programme, PRECARN through the Precarn Scholar programme and MITACS through the Accelerate BC programme.  xviii  This thesis is dedicated to Liz, Santiago and Andr´es. I love you all very much.  xix  Chapter 1  Introduction Ultrasound (US) imaging is a widely used tool in modern diagnostic procedures. It is a fast and safe method for acquiring data, as images can be viewed immediately and ionizing radiation is not involved. Ultrasound systems are portable, enabling them to be shared, have few special logistical requirements and are compatible with existing operating room and other hospital equipment. Additionally, most procedures are non-invasive and require no more effort than placing a probe on the patient’s skin. Ultrasound is therefore a tool of choice in many modern screening and diagnostic applications. One such application is the detection of deep vein thrombosis (DVT). DVT is a condition where blood clots or thrombi are formed within the deep venous system of the limbs. These clots may partially or completely occlude blood flow, or possibly break off from the vessel walls to which they are attached and flow proximally to the lungs where they can cause a pulmonary embolism (PE), which can be fatal. DVT is the third most common cardiovascular disease in the United States, after acute coronary syndromes and stroke [71], and it has been estimated that the incidence of clinically recognized in the general population is two cases per 1000 per year [55]. This proportion greatly increases among higher risk patients such as those who have recently undergone hip or knee replacement surgery. Some sources estimate that DVT affects approximately 2 million individuals each year in the United States alone. Recent estimates of deaths each year because of PE are 150,000 - 200,000 in the United States [125], and given the fact that approximately 80% of the emboli to the lungs arise from thrombi in the leg veins [57] detecting DVT in it’s early stages is clearly a priority. While several methods exist for detecting DVT, ultrasound is used as the primary imaging method to detect clinically suspected DVT. Thrombi are detected by performing a compression ultrasound examination (CUS). In a CUS exam, an examiner locates the vein of interest and applies gentle US transducer pressure while imaging the vessel. The vessel lumen will collapse unless there is a thrombus inside [30]. The loss of compressibility of a thrombus filled vein under gentle probe pressure has been identified as the most accurate, simple and useful diagnostic criterion for the diagnosis of DVT [8]. The examiner will scan the complete limb, performing the compression test at 2 to 5 cm intervals. The CUS can be augmented by the use of Doppler ultrasound and color flow imaging. There are however known limitations of the CUS examinations. It is known that sonography is unreliable for detecting asymptomatic or non occlusive DVT [34, 113], possible scan times of up to 40 minutes per limb, and that operator experience and expertise have a direct impact on the sensitivity and specificity of ultrasonography, where more experience has been shown to correlate with better detection, among other known issues. 1  In part because of these limitations, compression ultrasound is not typically used as a screening test. The benefits of having a screening test for DVT include the detection and treatment of asymptomatic but potentially life-threatening clots and the potential to reduce the number of patients treated (or ‘over medicated’) with prophylactic agents, which would be beneficial to all. A screening system that detects DVT would be of great use.  1.1  Motivation  It is evident that the proper detection of DVT is paramount for adequately treating this disease and avoiding the progression to PE, which could have very serious consequences. While current detection method are very good, certain issues remain. Increasing attention is being paid on obtaining 100% sensitivity and specificity rates, or in other words, on getting the best results possible without any false negatives or false positives. In addition, approaches are sought for adequately managing patients with suspected DVT or at high risk for DVT and patient treatments that improve the outcome for patients and reduce overall costs at the same time so as to make the detection and treatment of DVT and PE as efficient as possible. A ‘single ultrasound’ examination strategy for DVT detection has been suggested by some [27, 114, 141] which addresses these issues. Schellong et al [114] present a methodology that includes the strict standardization of the ultrasound examination protocol and a sound training of the sonographer. The authors argue that by making the examination more objective – through the standardization of the examination – and by reducing user variability through sound ultrasonographer training, the diagnostic workup of patients can be reduced to a single examination. One way to make the examination more objective is to develop a repeatable, measurement-oriented approach to quantify DVT, which is ultimately the approach that we have taken. If through the implementation of such a system user variability can also be reduced, then the barriers for a ‘scan everyone’ approach begin to disappear. The development of a system that can detect DVT using objective measures calculated from ultrasound images presents several interesting challenges. While real-time images of tissue and structures are readily available with ultrasound, boundaries may not always be clearly defined, characteristic ultrasound speckle may be present, and artifacts such as shadowing may appear in the images. Because of these factors, segmentation of features in ultrasound images becomes a difficult task. Appropriate methods, such as those that incorporate knowledge about the shape or pose of the desired feature must be used. Since detecting DVT using ultrasound is inherently dynamic, we cannot identify DVT based on a single image, methods are required for locating and tracking a region of interest over several images. In addition, measurements of the actual applied force will assist in the calculation of objective measures. A standardized, repeatable protocol for scanning for DVT is also necessary for a standardized examination. By defining and including specific tasks to be performed, a measure of the completeness of a scan can be generated. If a system is developed with this protocol in mind, this can ultimately have the effect of reducing user variability if all examiners have to perform the 2  same actions. The necessary training for users to become efficient with such a system may also be reduced. Such a DVT screening system has been outlined in the author’s previous work [60, 66], which is based on the principles of the compression ultrasound examination. In this system, ultrasound images are processed to obtain an estimate of the transverse vessel area with a contour detection algorithm, and the force applied by the examiner is measured. The area and measured force for each image during a CUS exam are used to generate a minimum to maximum area ratio and a vessel ‘stiffness’, using the single slope of the vessel area versus applied force, as measures that indicate the likelihood of DVT. A 3–D vessel representation of the scanned vessel is constructed using the detected contours and location measurements provided by a sensor mounted on the ultrasound probe. The compressibility measures are displayed as colours on the 3–D vessel model, and the completeness of the scan or adequacy can also be displayed. This prototype system had its limitations. Contour detection was limited to B-mode images and needed improvement. Seed point tracking was done using a very simple model, and advantage was not taken of the available location measurements. In addition, issues such as detecting vessel bifurcations had not been addressed. The vessel assessment and corresponding protocol was very basic, and the protocol was not properly integrated into the system. Only vessel compressibility was used to assess a vessel. Some of the main components of the user interface, such as the 3– D model display, the control window and the ultrasound display window, were present but were limited. In addition, the system was implemented on several computers linked across an internal network, limiting the frame rate and portability. The improvements to this system, interface and scanning protocol, as well as in depth validation of the resulting system(s) in both laboratory and clinical settings is one of the main topics of the current thesis.  1.2  Research Goals and Objectives  Several things are necessary in order to implement a reliable DVT screening or detection system that uses objective or numerical measures to determine the possibility of DVT within a vessel. Some of the main components are listed below. Feature detection algorithms must be developed and implemented that can provide the location of the vessel lumen in the ultrasound image given minimal input in a fast and accurate manner. This includes using the information from the various ultrasound modalities used to conventionally detect DVT to further enhance vessel detection. Feature detection should be fast enough to perform in real-time. The vessel of interest must be appropriately represented within the system, both to follow and track the location of the vessel in the ultrasound data, and to provide adequate visual feedback about the location being scanned. In addition, displaying the data generated by the system should be done in a straightforward manner to facilitate interpretation. Using any and all data available, measures that relate to the possibility of thrombus within a vessel must be generated. This includes, but is not limited to, determining vessel compressibility and blood flow properties within the scanned vessel. 3  The system must be easy to use by the target user population. Health care professionals such as nurses, ultrasound technicians or family doctors should be able to quickly and easily grasps the concepts behind the system and perform examinations using such a system. One of the aims of this system is to improve DVT detection by using objective measures to help standardize the examination protocol. As stated, this standardization has been suggested by several authors as a key factor for improving detection in ‘single ultrasound’ examinations and could allow examiners of varying skill to obtain accurate results and increase the availability of DVT screening. Specific research goals for the proposed DVT system are outlined here. • To develop, implement and validate contour detection algorithms that can detect vessels in ultrasound images, specifically in ultrasound images of the deep veins of the lower limbs. In addition, further requirements include: – The algorithm(s) should be able to perform in real-time. – The algorithm(s) should be able to deal with changing aspect ratios of vessels under different degrees of compression. – The algorithm(s) should provide a measure of the transverse vessel cross sectional area. – The algorithm(s) should be able to take input measurements from multiple sources, for example B-mode images, elastographs, or other data. • To develop, implement and validate tracking methods that can estimate the location of a feature throughout an image sequence. Specifically, the location of the vessel of interest in ultrasound images of deep veins should be determined. The method must perform in realtime. In addition, vessel tracking should be able to deal with situations that are common during ultrasound leg scans, such as detecting vessel bifurcations, automatic reinitialization if tracking is lost, and tracking under different conditions, for example when scanning or when performing a vessel compression. • To develop, implement and validate numerical measures that will indicate the possibility of DVT within a vessel segment. With respect to the compressibility measures proposed in [60], further validation and refinement is required. In addition, other measures that can help determine the possibility of DVT within a vessel are desired. • To develop, implement and validate an adequate user interface for the proposed system. Requirements for the interface include the display of collected ultrasound images, a representation of the scanned model, a control interface, and appropriate display of the calculated DVT measures. • To integrate all system components into a functional prototype, which can be validated through laboratory and clinical tests. It is expected that at the conclusion of this project, the system will have some clinical evaluation.  4  1.2.1  Functional Specification of a DVT Screening System  This Section outlines some of the functional specifications for a DVT screening system. Since we are modeling our approach on the conventional compression ultrasound examination, such a system should be able to identify the vessel cross-section in the ultrasound images in real-time, and should be able to measure the force with which a user is pressing down and measure the location at which this is being done. Properties of the vessel should be determined, including vessel compressibility, and the possibility of DVT within the examined vessel should be based on the measured data. A user interface should display a virtual 3-D model of the examined vessel in real-time and the examination results should be available ‘immediately’1 on the user interface. Ultrasound Imaging Unit The ultrasound imaging system must be capable of generating high quality B-mode ultrasound images, preferably using a linear element probe, at depths up to about 10 cm, with an image resolution of no less than 0.5 mm per pixel. Vessel features in the ultrasound image should not be less than about 20 pixels in diameter2 . This corresponds to using probes with frequency ranging from 4-12 MHz. Image resolution should be at least 8 bits per channel in a lossless format, and there should be minimal to zero delay for acquisition. Frame rate will vary depending on imaging depth and probe frequency, but should not be less than about 15 frames per second. In addition, typical ultrasound controls (image brightness, TGC (time gain control), depth, focus, etc.) should be included in the machine. The preferred embodiment comprises a single computer in which all software and hardware components for the ultrasound machine are integrated, as are the necessary components of the DVT screening system. This includes enough disk space, memory, display capabilities and network capabilities to meet these ongoing requirements. By including all subsystems on a single machine, the possibility of incorporating the complete system into a single software program increases. This can improve system performance and better approximate a final usable product. Sensors and Data Acquisition Any sensors should be suitable for use in a clinical environment. They should not interfere with conventional medical equipment or each other, be easily cleaned, and accept interference from conventional medical equipment and other factors in the environment. The sensors must be small enough to be mounted on an ultrasound probe. The force sensor must have at least one degree-of-freedom (DOF), specifically along the axis of applied force parallel to the image depth axis, with a suggested range of 0 to 100 N, with a resolution of at least 0.1 N. The data acquisition rate should be no less than the ultrasound image frame rate. 1 The use here of real-time and ‘immediately’ refer to how humans perceive a process. In this case, as long as the system does not delay the user from carrying out the examination protocol, the system will be considered real-time. 2 Resolution will be dependent on imaging depth. Assuming a vessel feature of about 10 mm, the feature will be approximately 20 pixels in size at 0.5 mm/pixel.  5  The location sensor must provide 6 DOF measurements, with a resolution of about 1 mm for position measurements and 1 degree for angular measurements. The acquisition rate must be no less than the ultrasound image frame rate, and preferably greater than 20 Hz for display purposes. The sensors should work throughout the work area (approx 600 mm x 600 mm x 450 mm when scanning a limb) and the sensors and system should be able to deal with problems such as line of sight (optical locations sensors) and electromagnetic interference (EM sensors). Data Processing Block The data processing block should integrate all the data acquisition, processing and display into one block. Specifically the main functions of this block are to acquire all data (ultrasound images, sensor data, user input), detect and track the vessel contours over time, calculate the compressibility measures and other vessel characteristics, and display the results to the users. Each of these procedures requires substantial computation and resources, and all procedures should be completed at the same rate (or better) as the ultrasound frame rate. Vessel detection and tracking should not take longer than half the available processing time for a single frame, determined by the ultrasound image and display frame rate. The calculation of the measures of vessel characteristics should be completed once all necessary data has been collected, without a noticeable lag by the user. Data display should appear fluid to the user, with rendered images at about 20 Hz or higher. The user should be able to readily interpret displayed data. Finally, all collected and generated data should be properly stored for review or audit.  1.3  Thesis Outline  This Chapter introduces the reader to the subject matter, and outlines the thesis goals. An overview of current literature, covering deep vein thrombosis and it’s detection, medical image segmentation and systems, and usability is presented in Chapter 2. The main system components, as well as system integration and user interface, are described in Chapter 3, while experimental results and validation of these components are presented in Chapter 4. Finally, a discussion about the proposed system and it’s possible impact, an overview of the main contributions of this dissertation and conclusions are presented in Chapter 5. Additional material includes a detailed overview of Kalman filtering in Appendix A, additional research on contour detection and vessel tracking is reported in Appendix B, a cost estimate of a DVT Screening System and the implications of such a system are discussed in Appendix C, details on the usability and system evaluations are presented in Appendix D and finally Appendix E includes required copies of all UBC Research Ethics Board Certificates of Approval.  6  Chapter 2  Literature Review This Chapter presents an overview of the different disciplines and areas needed to address the stated research goals and objectives, in order to develop an adequate screening system for deep vein thrombosis. An overview of deep vein thrombosis is presented in Section 2.1, followed by DVT detection methods in Section 2.2. This is followed by an overview of the different areas that comprise the subsystems of a proposed DVT screening system including image processing (Section 2.3.1), force and location sensing (Section 2.3.2), 3-D modeling (Section 2.3.3) and characterization of physical properties (Section 2.3.4). Finally, the topic of system usability is presented in Section 2.4.  2.1  Deep Venous Thrombosis (DVT)  There is no doubt that deep venous thrombosis (DVT) of the lower limbs is a common and dangerous condition, with the potential to lead to fatal pulmonary embolism. Despite advances in primary prophylaxis, venous thromboembolism still occurs in a considerable number of high-risk surgical patients [41, 87], and suspicion of DVT is a common reason for medical referral to hospital [108]. Deep venous thrombosis is a vascular disease that mostly affects the deep veins of the lower limbs. Thrombi are intravascular deposits usually formed in regions of low or disturbed flow, such as large venous sinuses or valve cusp pockets as shown in Figure 2.1. They develop from the components of blood, which are mainly fibrin and red cells and with a variable platelet and leukocyte component [29, 71]. Approximately 40% will resolve spontaneously, 40% will become organized, and 20% will propagate. Although there are many factors that can contribute to the development of DVT, such as stasis, endothelial injury, and the condition of the coagulation system [29], it has been reported that the three main risk factors for thrombus formation are age greater than 75 years, previous history of deep venous thrombosis, and underlying malignancy [6]. Thrombi can partially or completely obstruct the flow of blood through venous pathways and often break off to form pulmonary emboli, which pass to and obstruct the vessels of the lungs and is frequently fatal. Alternatively, if clots do not break off, in the long-term they can cause venous insufficiency, where venous valves become damaged permitting reversed blood flow. DVT can extend throughout the deep venous system (illustrated in Figure 2.2), including the iliac veins, the common femoral (CFV), the femoral (wrongly but universally named the “superficial” femoral) (SFV) and deep femoral or profunda femoris (PROF) veins, popliteal veins (POP), the deep veins of the calf (including the anterior (AT) and posterior (PT) tibial veins, and peroneal (PER) vein), 7  Figure 2.1: Venous valves shown in a closed position when pockets are formed behind the cusps, with normal venous flow (blue) and blocked return flow (green). and can also involve the superficial leg veins, such as the long saphenous vein (LSV) [71]. It has been estimated that the incidence of clinically recognized DVT in the general population is two cases per 1000 per year [55]. Other sources estimate that DVT affects approximately 2 million individuals each year in the United States alone. Of these, approximately 600,000 patients develop pulmonary embolism (PE) [71, 134], and recent estimates of deaths each year because of PE are 150,000 - 200,000 in the United States [125]. DVT is the third most common cardiovascular disease in the United States, after acute coronary syndromes and stroke [71]. Recurrent DVT must also be considered, as it has been reported at about 5% after 3 months, about 5-10% after 1 year, and about 30% after 8 years [71]. The patient’s risk for DVT can be classified at different levels. High risk patients usually include the elderly, bedridden, post-surgery, and post-trauma. Patients who have undergone hip replacement typify a high-risk group because of their advanced age, trauma, and surgical state [29]. There is a 20% - 30% chance that DVT will occur in patients who have undergone recent hip replacement surgery [27], or about a 30% chance if patients have undergone moderate surgery and receive no perioperative antithrombotic prophylaxis [134]. Schmidt et al [115] also report thrombosis in up to 30% of patients after orthopedic surgery, and Ilahi et al [75] report an overall DVT rate of 9.9% and a proximal DVT rate of 2.1% after knee arthroscopy in patients without antithrombosis prophylaxis. Most postoperative thrombi arise in the calves, and approximately 40% - 50% of patients with skeletal trauma present with calf vein thrombosis. PE does not occur without DVT, and most deaths due to PE occur within a few hours of its onset [29]. Approximately 80% of the emboli to the lungs arise from thrombi in the leg veins [57]. Among patients with clinically confirmed symptomatic PE, DVT is found in about 70%, while among patients diagnosed with DVT, pulmonary emboli are detected in about 50% [71]. The spontaneous course of the disease is quite unfavorable, since mortality due PE in untreated DVT 8  Figure 2.2: Diagram showing the major vessels of the deep venous system. patients exceeds 20% [59]. In up to 30% of episodes of PE, the outcome is death [17]. It has been suggested that prevention of PE must be the goal, and this is achieved only by controlling and preventing DVT [29]. One solution is the practice of screening patients postoperatively to detect DVT in its early stages, as it reduces the occurrences of major PE because of earlier diagnosis and treatment of DVT [134]. In this setting, a fast, simple and reliable test that could be carried out directly where the patients are located would be extremely useful [55]. A recent study [121] has also found evidence of a link between DVT and acute cardiovascular events such as heart attacks and strokes, with an increased risk of the latter in patients diagnosed with the former. The increase in the risk was most pronounced during the first year after the DVT (up to about 90% increase), and were similar in patients both with ‘provoked’ thromboembolism cases (diagnosed malignancy before or within 90 days of DVT, or fracture, surgery, trauma or pregnancy in the 90 days before DVT) and ‘unprovoked’ DVT (the remainder of the patients with DVT). Although the association is still not entirely clear, DVT screening may help identify patients with a high risk of other cardiovascular diseases. An incorrect diagnosis of DVT can have very serious consequences. The complete spontaneous 9  lysis or gradual subsiding of the symptoms of large venous thrombi is uncommon, and even when patients receive treatment there is a less than 10% chance that complete lysis will occur [71]. A false-negative diagnosis would therefore subject the patient to the risks of clot propagation, embolization and death. An expert panel estimated the probability of developing post-thrombotic syndrome to be ∼ 33% in untreated patients [56]. A false-positive diagnosis would subject the patient to needless anticoagulation therapy and its risks [29]. Proper detection is therefore extremely important.  2.2  DVT Detection Methods  The multiple factors that can contribute to the formation of DVT are sufficiently non-specific that it is impossible to definitely identify an individual who will develop DVT [29]. Moreover, once an individual presents with typical symptoms for DVT, a long list of differential diagnosis must be considered before settling for DVT, which include popliteal venous aneurysm, popliteal artery aneurysm, pseudoaneurysms, superficial thrombophlebitis, popliteal (Baker) cyst, cellulitis, enlarged lymph nodes extrinsically compressing veins, heterotopic ossification, hematoma, postphlebitic syndrome, ruptured muscle or tendon, muscle strain, internal derangement of the knee, cutaneous vasculitis, lymphedema, and chronic venous insufficiency [17, 29, 71]. It has been well documented that the diagnosis of DVT based on clinical signs and symptoms alone is inaccurate and unreliable [26, 28, 29, 33, 34, 37, 48, 49, 58, 72, 108, 122, 137]. Although some prevalent symptoms such as swelling in the lower limb, and occurrence of multiple symptoms have been reported for patients with positive DVT results [37], fewer than a third of patients actually present with classic DVT symptoms [48, 134]. Of patients with proven PE, less than 20% have signs and symptoms suggestive of DVT [28, 71], but of these about 70% have evidence of DVT [71]. A study by Cranley et al [28] reports that only 4% of fatal cases of PE presented warning symptoms of thrombophlebitis and only 10.6% of patients were diagnosed with peripheral venous thrombosis prior to fatal PE. Additionally, only 9% of patients with diagnosed PE had positive leg signs. Another study by Cronan [29] reports that for patients who died while in hospital, 10% died of PE over 5 years. Of these, 83% had DVT at autopsy, but only 19% had symptoms before death. One exception to clinical evaluation of DVT is the Well’s score. The Wells’ criteria [135, 137] have generally been seen as the most reliable clinical score developed, having been both derived and then validated in patients with suspected DVT [19]. This set of criteria consists of nine clinical features and allows the clinician to obtain a score that predicts the pretest probability of the presence of a DVT. The assessment of a patient’s pretest risk for DVT is increasingly being used in conjunction with objective testing to obtain a more reliable measure[19, 53]. A reliable clinical prediction index can help to reduce the number of suspected DVT patients (up to 50%[93]) and the cost to both the payer and provider alike [19]. The nine Well’s features are the presence of active cancer (+1), paralysis or a cast (+1), recent immobilization or surgery (+1), tenderness along deep veins (+1), swelling of the entire leg (+1), a difference in calf circumference of more than 3 cm (+1), pitting edema (+1), the presence of 10  collateral superficial veins (+1), and whether an alternative diagnosis is likely (–2). Of these features, the determination of the likelihood of an alternative diagnosis is the least objective and yet the classification is highly dependent on this. Some have suggested removing the possibility of an alternative diagnosis in order to make the Well’s score more objective [94]. A patient with a Well’s score higher than 3 is classified as high risk, a score between 1 and 2 is classified as medium risk and a score of 0 or less is classified as low risk. However, Well’s score cannot be used in patients with a prior DVT (up to 16% of patients) [19]. Because of the unreliable nature of clinical signs, the diagnosis of DVT must be based on objective paraclinical tests [85]. In order to assess the reliability of a detection method, four indicators are used. Sensitivity (ability to correctly identify positives) is the proportion of true positives over all positives and specificity (ability to correctly identify negatives) is the proportion of true negatives over all negatives. The positive predictive value is the proportion of correct positives, and the negative predictive value is the proportion of correct negatives. It is important that these examinations be sensitive to correctly identify diseased individuals, and specific to correctly discard healthy individuals [85]. Common clinical tests for detecting DVT include ultrasound (US) with its different modalities (including compression ultrasound (CUS) [30], and duplex ultrasonography (DUS), which combines duplex ultrasound and spectral and/or colour Doppler [53, 111]), and contrast venography (CV), CT or MRI venogram, and the laboratory test D-Dimer [6]. Other examples of techniques that add diagnostic value include strain gauge plethysmography (SGP) [55, 108] and light reflection rheography (LRR) [122], however their diagnostic performance is inadequate to act as a standalone test in DVT diagnosis [89]. The D-Dimer study is being used more frequently as a screening test with 99% sensitivity in detecting thrombus, but specificity is only approximately 50% [6]. Contrast venography is still considered the most accurate diagnostic method for DVT in asymptomatic patients [49]. However, duplex and colour Doppler sonography (DUS) is currently the technique of choice over contrast venography [5, 34] for the diagnosis of DVT in symptomatic patients, because it has proven safe and cost-effective, with a very high sensitivity and specificity (96% and 98%, respectively) for the diagnosis of proximal DVT [53]. Some of the major advantages of ultrasound over venography include the fact that US is noninvasive, painless, requires no radiation exposure, and can be done at the bedside if necessary [113].  2.2.1  Contrast Venography (CV)  In contrast venography, contrast medium is injected into a vein through the patient’s foot, and it is permitted to flow through the patient’s venous system. After a short period of time, radiographic images are taken of the filling of the veins by the contrast medium. Healthy veins will fill up completely and show up on the radiographic images, but diseased veins will not. By correctly interpreting this absence of filling of the veins, DVT can be diagnosed. Nonetheless, CV is not considered a suitable screening test because of its invasive nature [49], as well as because of the use of ionizing radiation, its high cost, it is time consuming, and because it has a significant rate of inadequate examinations and is subject to misinterpretation [141]. Additionally, venography has a 2% risk of causing DVT itself, and it is uncomfortable for the 11  (a) Uncompressed SFV and SFA (top)  (b) Partially compressed SFV (bottom)  (c) Compressed SFV, with SFA (top)  Figure 2.3: Image sequence showing the compressibility of a SFV (bottom) as force is applied (left to right) with an ultrasound probe.The artery remains with a similar area throughout (top). patient [141].  2.2.2  Ultrasound Imaging  Currently, sonography is the primary imaging method to detect clinically suspected DVT, having replaced contrast venography as the standard diagnostic test [49, 59, 94, 111, 125]. Ultrasound systems are portable, enabling them to be shared, have few special logistical requirements and are compatible with existing operating room and other hospital equipment. On ultrasound a fresh thrombus (< 7-10 days old) is mostly hypoechoic, homogeneous, partially compressible, and sometimes floating, and the vein diameter is enlarged. On the contrary, an older, organized thrombus is hyperechoic, heterogeneous, incompressible, and firmly adherent to the vein walls, and moderate or no vein enlargement is visible [53].Viola et al [130] suggest that an increase in echogenicity associated with the formation of thrombus is mostly due to ultrasonic backscatter and changes in the speed of sound and attenuation. The objective of a compression ultrasound exam is to detect DVT in the lower limbs by observing the transverse area of the veins when gentle pressure is applied to determine if the venous system is obstructed. Loss of compressibility of a thrombus filled vein under gentle probe pressure, i.e. CUS, is the most accurate, simple and useful diagnostic criterion for the diagnosis of DVT [8]. Figure 2.3 shows the compression of a healthy SFV and accompanying superficial femoral artery (SFA). This lack of venous compressibility accurately indicates high probability of anechoic thrombus. If a vein does completely collapse, the possibility of DVT in that section of the vein is, on the contrary, very small. Further information obtained from an ultrasound examination includes thrombus extent (mainly 12  its upper limit) and characterization (fresh or organized, free-floating or attached, and partially or totally occlusive) [53]. When colour is used, information of the absence, reduction, or turbulence in venous flow can be used to reinforce the diagnosis. In addition, Andrews et al [6] report that two negative ultrasound studies a week apart exclude the diagnosis of DVT. The sensitivity and specificity for B-mode and duplex sonography has been widely reported, with consistent results throughout various studies. Overall values for US specificity and sensitivity are reported at 98% and 95%, respectively, when compared to venography [29]. The sensitivity of US for calf vein thrombosis is 73% [49]. However, it is well known that sonography is more accurate in detecting DVT in symptomatic patients than in asymptomatic patients [58]. Pooled results from six studies showed an overall sensitivity for the detection of proximal DVT of only 59%, although specificity was 98% [134]. Also, compression US has been reported as being much less sensitive for diagnosing DVT following high-risk surgical procedures, such as total hip or knee arthroplasty, where researchers obtained a sensitivity of 62% for asymptomatic patients [49]. The sensitivity and specificity for colour Doppler or flow imaging (CDI or CDFI) is known to be worse than for CUS or duplex US. Some have reported overall sensitivities of 80% - 90%, but in most reports the sensitivity of CDI is below 80% for proximal DVT, and quite lower for calf DVT [85]. The trials with higher sensitivities were conducted by experienced persons, suggesting that operator experience can affect the outcome considerably. One such study by Becker et al [14] presents results for CDI with a sensitivity of 98% and specificity of 96%, and in all exams where DVT could not be recognized, patients either had thrombosis in the calf veins or isolated partial thrombosis of an iliac vein, both difficult cases to identify with CDI. False positives were also present in the same anatomical areas. Lower sensitivity was reported by Fern´andez-Canton et al [48] in their study, with an overall sensitivity of 53%, specificity 100%, positive predictive value 100%, and negative predictive value 68%, while for the calf sensitivity was 40%, specificity 100%, positive predictive value 100%, and negative predictive value 79%. Because of low sensitivity, CDI cannot stand alone as a screening method for DVT [85]. Examination Procedures Usually an CUS examination comprises examining the deep venous system beginning at the common femoral vein (CFV) at the inguinal ligament, and proceeding distally until reaching the trifurcation of the calf veins [137]. The veins are located and compressed while imaging on a transverse plane at locations typically between 1 to 5 cm apart. Care should be taken when performing an examination because of the common possibility (18%) [109] of duplicated venous segments [17]. This may pose a problem if a thrombus within one limb of the duplicated segment is not visualized and compression of the non-thrombosed duplicated segment is interpreted as a negative examination [49]. In symptomatic patients, thrombosis usually involves whole or multiple venous segments, and it is adequate to scan every 1-2 cm [29]. Examinations may take about 20-30 minutes per extremity, but examination times of up to 40 minutes per limb have been reported in the literature [110]. The main criterion for diagnosing positive DVT when using CUS is the lack of compressibility 13  when gentle probe pressure is applied. Several additional findings may be observed including an expansion of the venous lumen for acute DVT, absence of blood flow, and loss of phasicity (regular flow variations modulated by patient’s breathing) or augmentation (increased flow when distal compression is applied). Colour imaging can also be used to assess for some of these criteria, as a colour void or filling defect within a vessel lumen with DVT would be observed [14, 17, 29, 49, 85], where a healthy vein would fill with colour. Colour ability can also be used to identify vein from artery [48]. Thrombi are usually hypoechoic, although this varies depending on transducer frequency and age of clot. Clot echogenicity is sufficiently unreliable to diagnose acute clot or to assess the age of the clot [29]. With time the clot will usually become echoic [49]. If a 5 MHz transducer is used, vessels ranging in size from 1 mm to slightly over 1 cm can be clearly visualized over a depth of 6 cm [17]. When performing a compression examination, care should be taken not to deform the arterial wall, as this indicates that excessive pressure is being used [29]. The area of the inguinal ligament and adductor canal are difficult or impossible to compress, in which case two handed compression can be used, using the second hand to compress the tissue from below. Normal Doppler findings in these areas can also be used to demonstrate a healthy venous segment [29]. Another test that aids in the diagnosis of DVT in proximal deep veins of the lower leg [29] is the Valsalva Maneuver, or to have the patient bear down. The normal physiological response of the femoral vein to the this procedure is a 50% to 200% increase in diameter. If venous thrombi are present, the vessel has a limited or absent response and does not increase in caliber. Duplex Doppler and CDI are helpful to identify veins but are not mandatory [49]. When attempting to visualize the deep veins of the legs, especially with patients with leg swelling with pain or edema, a combination of modalities may be quite useful, including CUS, augmentation, colour and power imaging [17, 141]. Power Doppler depicts the density of blood cells rather than the velocity of blood cells, and is suited to multidirectional flow detection because of its relative angle-independence [138]. Additionally, power Doppler is more sensitive in detecting venous blood flow, and is less degraded by noise and clutter [38], although it does not indicate directionality of flow [17]. If DVT is found, the proximal extent of the clot should be determined [29]. Proper reports are critical for documenting the extent and location of the DVT. These should indicate the maximal compressed diameters of the examined proximal vessels [49]. Examination of femoro-popliteal segment is quick and easy [48], but it is possible for small, non occlusive thrombi to be overlooked with duplex US [109]. The examination of the distal system on the other hand can be tedious, not as sensitive, and the use of all available imaging modalities can assist in thrombus detection. Having experienced ultrasonographers with good technique and a strictly defined protocol is extremely important, as this can affect the accuracy and completeness of the examination [17, 48, 59, 106]. Several authors have suggested a ‘single ultrasound’ examination strategy for the diagnosis of DVT [27, 114, 141]. The most recent presented by Schellong et al [114] is a methodology which includes two important prerequisites. The first is the strict standardization of the ultrasound examination protocol, which includes restriction to B-mode ultrasound, exact definition of all venous segments, and having the patient in the sitting position when examining segments below 14  the knee. The second prerequisite is a sound training of the sonographer. The authors argue that by making the examination more objective (standardization of the examination) and by reducing user variability (sound sonographer training), the diagnostic workup of patients can be reduced to a single exam. Several courses of action have been suggested once an outcome is available. If the outcome is positive, no further study is needed [48] and appropriate treatment can start. On the other hand, if the outcome is negative there are several options. If the examination included careful evaluation of the calf veins and the patient was symptomatic and without risk factors for DVT, the disease can be excluded [27]. Others have suggested that the outcome of the exam should be combined with the patient’s pretest probability for DVT, which could improve the diagnostic process and decrease the overall cost [137]. If there is ongoing high clinical suspicion of DVT with a negative US exam outcome, a repeat scan should be considered especially if the calf veins were not properly visualized. This is based on the potential risk of missing calf DVT which may propagate proximally [141]. A patient’s history of DVT should also be considered. A normal exam can confidently exclude acute DVT, but an abnormal result could not only suggest acute DVT, but chronic changes as well [32]. Adverse outcomes after negative DVT studies, such as PE and thigh DVT (thromboembolic complications), have been reported among postsurgical patients [58], or among moderate or high risk, hospitalized patients [141]. Poor clot detection in orthopaedic patients may be because of the presence of calf clots, and small non-occlusive clots in the larger veins [29]. Calf DVT There is much controversy as to the risk of calf thrombosis, and how detailed a study should be done on the calves to exclude DVT. Because of small vein caliber and varied anatomic locations, calf DVT exams are a time-consuming process [49]. It is estimated that an additional 15 - 20 minutes per leg is needed to perform these exams [29]. Even so, some investigators believe the examination of calf vessels is a useful procedure [8, 49] because of the risk of thrombus propagation. In general, 10% - 20% of patients with symptomatic DVT will have isolated calf thrombosis, and 20% - 30% of symptomatic and asymptomatic calf vein DVT will eventually extend to the proximal venous system [29, 49, 71, 82, 134]. This increases the risk of clinically important PE [49] to 40% - 50% [82]. Of calf clots, about 40% remain isolated, while another 40% quickly lyse (symptoms gradually subside) [29]. Additional caution should be exercised when evaluating the calves because of the consistently high number of indeterminate studies [57, 125]. Nonetheless, the direct evaluation of calf veins is successful in 60% - 90% of patients [29], and experienced operators can often achieve visualization of the proximal veins of the calf [71]. The sitting position is extremely useful, as it causes passive distension of the calf veins, which permits them to be more easily visualized. A weak point exhibited in some calf DVT detection studies is that flow detection with Doppler tends to be more difficult in peripheral veins [14], which should come as no surprise because of the lower flow velocity. Augmentation or the use of a signal enhancing agent can improve vessel visualization [106]. 15  Chronic DVT If thrombus material remains inside the vessel lumen, even after detection and treatment, chronic DVT may develop causing chronic venous insufficiency, exemplified by the postphlebotic syndrome (symptoms of swelling and localized pain). Fifty-three percent of patients with documented DVT experienced the postphlebotic syndrome [32]. Over a period of weeks to months, fibrous tissue or thrombus material accumulates within the clot and the clot retracts and remains as organized thrombus along the vein wall, becoming increasingly dense and echogenic, and making the wall thicker and more resistive to compression [29, 32]. When the vein is compressed the new smaller lumen collapses, but the original vein walls remain separated, giving the impression of an uncompressed vein, such as would happen with an acute clot [29]. The distinction between chronic and acute clot can therefore be difficult [17] and remains a problematic area. The use of CUS is much less specific for recurrent or chronic DVT, and may not be a reliable follow-up study after an episode of acute DVT [32]. Indicators for identifying recurrent or chronic DVT with ultrasound include areas of thrombosis not observed in the initial scan, or considerable enlargement (> 2 mm) of the compressed vein diameters [49]. With acute thrombus, the vein enlarges often to twice the size of the accompanying artery, while in chronic disease, the vein is similar in size to the adjacent artery [17, 29]. In addition, collateral veins are an indication that chronic disease is present [29, 32]. Wall thickness can be assessed with CDI [29]. When thrombus material is deposited and accumulated on venous valves, these may be destroyed and chronic venous insufficiency syndrome ensues [17]. This syndrome can be identified if a reversal of blood flow is observed within a vein following augmentation [29]. Selective or Reduced Scanning Several clinical findings seem to support proposed selective scanning procedures. For example, it has been found that thrombi are rarely isolated to the iliac, femoral or popliteal regions (5%) but commonly isolated to deep veins of the calf (51%) [109]. This supports a routine scan of the calves but suggests that an in-depth exam of the iliac vessels may not be needed. On the contrary, results from a more recent study by Gottlieb et al [57] suggest that clinically significant calf thrombi are present only in patients with physical signs or symptoms in the calf at the site of the thrombus. The researchers for this 2003 study developed a protocol that does not routinely examine patients without calf symptoms or signs, and concluded that there is no increased risk of clinically significant propagation of thrombi into the thigh veins or PE. This can also be translated into a potential 10 minute time savings per leg [57]. The researchers do state though that it would be appropriate to evaluate the calf in patients with signs or symptoms. It has been suggested by Fraser et al [49] that an examination limited to the common femoral or popliteal vein regions may be an option as a screening test for the diagnosis of DVT in symptomatic patients. On the other hand, for patients that are suspected of having PE and have no symptoms or risk factors, a study by Sheiman et al [119] shows that screening with ultrasound is of no clinical utility, and should not be ordered as a first-line for PE. The need to perform bilateral examinations is another source of controversy. It has been suggested [118] that in patients with unilateral symptoms, US screening for DVT should be limited 16  to the symptomatic extremity regardless of predisposing factors, as this would decrease scanning time and cost without a decline in the DVT detection rate. [34] cites that previously, others had concluded that contralateral scanning is not clinically necessary in patients with unilateral symptoms, but their own findings suggest that both lower limbs should be systematically examined with ultrasound, in symptomatic patients as well as asymptomatic patients at high risk of DVT. In [53] the authors state that bilateral DUS examination of lower limb veins should be performed as the initial examination in the workflow of pulmonary embolism only in high-risk patients or when screening asymptomatic patients. An important and interesting finding of the study by Asbeutah et al [7] was an unexpectedly high incidence of venous reflux in the apparently unaffected limb. Although these non-DVT limbs were not investigated at presentation, the data is consistent with the hypothesis that DVT may result in a more systemic disorder of venous function. Screening vs. Diagnosis Many approaches for screening for DVT have been suggested and investigated. Arguments for and against routine screening have been put forth, and it seems the efficacy of screening may depend on the patient population being examined. Advocates argue that routine screening has several potential advantages including detection and treatment of asymptomatic but potentially life-threatening clots and the potential to reduce the number of patients treated with a prophylactic agent after discharge, thereby reducing the cost, inconvenience, and morbidity associated with thromboembolism prophylaxis. However, routine screening also has some disadvantages including imperfect accuracy of screening tests, the potential of screening to lead to overtreatment of patients who otherwise are asymptomatic, the potential undertreatment of patients with clots that are not detected, and the cost associated with routine screening [15]. Because screening altered management in 26% of patients in a published study [112], the authors report that screening duplex exams to detect lower extremity DVT in rehabilitation patients is useful. Others that support screening include Elias et al [41], because of the high accuracy and reliability of extended US, and screening at hospital discharge is suggested by Verlato et al [129] because of the relatively high incidence of proximal DVT in patients undergoing major orthopaedic surgery under prophylaxis. They elaborate that the negativity of this test has the potential of safely preventing the extension of anticoagulation beyond hospital stay. Others seem to be on the fence. Hirsh et al [72] state that screening for associated thrombophilia is not indicated routinely, but should be performed in selected patients whose clinical features suggest an underlying hypercoagulable state. However, many do not support routine screening, because findings support that it is sufficient to give patients adequate or extended prophylaxis without surveillance to protect against proximal DVT [15, 78, 84]. However, the situation is different for high-risk patients who have not received prophylaxis [84]. Others [115] have found that ultrasound screening for thrombosis does not reduce the rate of proximal thrombosis and symptomatic pulmonary embolism after hip or knee replacement surgery. The American Thoracic Society states that, for asymptomatic patients, the 17  sensitivity of US is too low for the technique to be considered reliable as a screening test, even in high risk patients [5]. Examples are cited in [53] stating that systematic Doppler ultrasound screening is not justified in asymptomatic patients with high prevalence of DVT (specifically ambulatory cancer patients undergoing chemotherapy), and the authors of a meta-analysis of studies with minimized bias concluded that the low sensitivity of venous sonographic imaging (reported at 54-70%) limited its usefulness as a screening test. When screening is performed, most authors agree that ‘reduced’ screenings should not be done [6], as diagnostic efficacy may be reduced [5]. In all cases, the ultrasound examination must be complete, from the ankle to the groin and abdomen (iliac veins and inferior vena cava), and bilateral. The wrong way would be to omit examination of the calf or the contralateral limb, thus reducing the clinical significance of negative findings [34]. However, an older paper [50] reports that DVT limited to a single vein occurs with sufficient frequency that the US screening survey cannot be abbreviated without loss of diagnostic efficacy. Zierler [144] has suggested that national standards in the form of clinical pathways using evidence from the literature need to be developed and utilized in order to achieve optimal utilization of the vascular laboratory. In addition, referring physicians need to be educated on the accuracy and limitations of duplex ultrasound scanning to avoid both overtreatment and undertreatment of acute DVT. Hybrid Screening Systems Some authors have investigated the use of ultrasound combined with other measurements as possible screening and diagnostic methods for DVT, that may overcome some of the shortfalls of scanning using only CUS examinations. Hirsh and Lee [72] suggest that making a diagnosis of DVT requires both clinical assessment and objective testing because the clinical features are nonspecific and investigations can be either falsely positive or negative. It has been proposed [81] that an effective means of excluding proximal DVT when screening outpatients is for patients with a difference in calf circumference of 2 cm or less and a negative Ddimer may undergo nonemergent DUS, and that patients with a positive D-dimer or asymmetrical calf swelling of more than 2 cm alone or in combination should undergo emergent DUS. A screening method for DVT and duplex ultrasonography using patient’s active maximum ankle dorsiflexion has been presented in [113]. Flow ratios of the peak flow signal with active maximum ankle dorsiflexion and that at rest was measured using duplex ultrasonography in the bilateral femoral veins. The authors report that the flow ratios with DVT were significantly lower than those without DVT. All of the thrombi in this study were detected in the operated lower limbs and were nonocclusive. They also suggest that, because of its simplicity, their measurement procedure is probably operator independent. Sebastian et al [117] assessed DVT by the quality of the flow signal (“whoosh”) using portable Doppler ultrasound, but concluded that this method does not have adequate accuracy to be used as a quick screening tool for DVT.  18  Costs Perone et al [102] reported that the cost for a ultrasound examination ranged from $90 - $300 USD, with a baseline value of $180 USD. When including the costs of the hospital room and treatment, a serial ultrasound examination (repeating a CUS 7 days after the initial exam, to rule out calf thrombus propagation) totaled $1,482 USD per patient. As presented above, several authors have suggested a logical solution to cost reduction: a single examination, which may include the use of signs and symptoms, that can detect DVT in one pass. If no repeat examinations are necessary, all costs associated with the repeat examination, including examination room, ultrasonographer wages, and other associated costs, can be reduced. Also, an extremely important benefit of an adequate single examination is the proper detection of patients with DVT. From a costs point of view, any costs associated with treatment of the disease at more advanced stages are reduced or saved if DVT is found in its early stages. Savings may also be obtained from an improved screening method that can exclude patients without DVT, and only refer patients to a complete US examination for confirmation, thereby decreasing patient load from not having all patients undergo complete US examinations. An increase in cost may arise from the need for experienced sonographers. If a DVT detection system relies on having experienced sonographers with a ‘sound training’ in order for the detection to be reliable, this may drive wages up for experienced sonographers. The reduced cost of a faster examination may compensate this but ideally the method for detection would be user-independent. Pitfalls of US Methods Some of the pitfalls of ultrasound methods have already been made evident. Sonography is unreliable for detecting asymptomatic or nonocclusive DVT [34, 113], and suffers from a lack of sensitivity at the calf level [15, 34, 94] when compared to venography, even when using Doppler imaging [139]. Also, isolated thrombi in the iliac and superficial femoral veins within the abductor canal are rare but difficult to detect, and therefore are easily overlooked [94]. It has been suggested that the sensitivity of DUS is greater in populations with a greater prevalence of DVT [53]. It has also been reported that accurate diagnosis of post-surgical DVT can be time-consuming [117]. However, it has been suggested [128] that duplex ultrasound screening for asymptomatic DVT is a suitable instrument for internal hospital quality control in thrombosis prophylaxis, and routine use can be recommended at least in high-risk patients, not only from the medicolegal aspect but also from the purely economic aspect. Operator experience is also a recurring theme. Many authors [5, 15, 53, 113, 139] have reported that operator experience and expertise have a direct impact on the sensitivity and specificity of ultrasonography, where more experience has been shown to correlate with greater accuracy. However, it has also been found that this widely advertised operator dependence of the sonographic diagnosis of DVT is not prohibitive, and good agreement (90%) between operators has been reported [34]. The lack of standards for US examinations may also affect the outcome [53]. Other pitfalls and limitations of venous DUS are related to difficulty imaging veins, and possibly vein duplication, because one vein may be normal while the other contains a thrombus [53]. Also included are technical limitations such as: a poorly compressible vein due to obesity or edema; an 19  echogenic lumen due to low velocity blood flow as in low cardiac output; and vein distention due to congestive heart failure. Blood flow echogenicity resulting from blood stasis and erythrocyte aggregation may be a significant source of false positive results. In these situations, dynamic tests (flow augmentation produced by passive limb raising or upstream muscle compression) may ensure a correct diagnosis [53]. The value of colour Doppler ultrasound is also somewhat suspect. It has been reported that augmentation rarely provides additional information in the diagnosis of DVT and the lack of usefulness and patient discomfort may justify removal of augmentation from the routine study [90]. Similar findings were reported in [87], where it was concluded that colour Doppler ultrasonography does not increase the detection rate for asymptomatic DVT over compression ultrasound, and it was reported in [117] that portable Doppler ultrasonography does not have adequate accuracy to be used as a screening tool for DVT. There is also the possibility that excessive compression during a CUS examination may dislodge a thrombus itself, causing a pulmonary embolism. Throughout the almost 20 years that this technique has been in use, no more than three cases of a thrombus dislodging during an examination have been reported in the medical community [45, 101, 116]. Fatalities were in no case reported.  2.2.3  DVT Screening System  A DVT screening system has been presented by Guerrero [60] and Guerrero et al [63, 64] that uses a contour detection algorithm to identify vessel contours in US images and generate an estimate of the transverse vessel area that, combined with force and location measurements, can be used to calculate objective measures of vessel compressibility in real-time. The system also generates a 3-D model of the geometry of the scanned vessel that can display the results of the compressibility measures mapped as colour. This system has been further developed and extended, and is the topic of the current dissertation. An overview of the system, as well as an in-depth validation of the compressibility measures and system performance were reported in [66], while a detailed report and validation of vessel segmentation and tracking was presented in [67]. In addition, the results of a pilot study using the extended system were presented in [65]. Details are presented in the following Chapters. System Hardware and Instrumentation A sensorized hand-held ultrasound probe consisting of a pair of nested shells, and force and location sensors, is used. Two aluminum shells surround a linear 9-4MHz ultrasound probe. The inner shell is fixed to the probe, while the outer shell is connected to the inner shell through the 6 degreeof-freedom (DOF) force/torque sensor (Nano25, ATI Industrial Automation, Inc.) at the rear, as shown in Figures 2.4(a) and 2.5. The examiner can grasp and manipulate the ultrasound probe in a normal manner, and all applied forces and torques can be measured. An electromagnetic sensor (PCIBird, Ascension Technology Corp.) is rigidly attached to the rear of the outer shell through a 5.33” plexiglas rod, and therefore the location of the image plane (and extracted 2-D contours) can be calculated using a calibrated transformation. 20  (a) Exploded View of Probe Shells  (b) Ultrasound Probe  Figure 2.4: Sensorized ultrasound probe used in the DVT screening system. Schematic (a) shows the inner and outer shells, and force /torque sensor, while (b) shows a user grasping the outer shell. The prototype system was implemented using two PCs, the first a PC-based ultrasound machine (Ergosonix RP 500, Ultrasonix Medical Corp., Burnaby, Canada) and a second conventional PC housing the force and locations sensors and system software. Image acquisition was done using the Ultrasonix libraries, so no additional image capture hardware was necessary. Contour Detection and Seed Tracking The transverse vessel area was estimated using the estimates of the semi–major and semi–minor ellipse axes generated using a star-Kalman algorithm with a 3–parameter ellipse model. A discrete contour is also generated when using this algorithm and data is pre-processed using a median filter. For each image, the contours are validated by calculating the radial error between the detected contour and the estimated ellipse. If the error was larger than a threshold, the search area was reduced and contour detection was repeated. Once a contour was validated, the estimated ellipse parameters are used as initial conditions for the next image. The seed location was tracked using a 2–D constant velocity model that estimates the location of the next seed point using the seed location and location measurements from the current frame. A drawback of this original system is that it was calibrated only for a single depth setting. Model Building Given a detected contour, points were transformed into the world frame, and used to construct a 3–D model of the scanned vessel using the transformed points as vertices. A custom triangulation method was used to construct the necessary polygons to build the model. Each point also has associated data that corresponds to compressibility measure and is mapped to the vessel model 21  Figure 2.5: Sensorized ultrasound probe with exposed inner shell, force–torque sensor casing, and location sensor mounted on a Plexiglas rod. surface as colour. Vessel Compression Assessment The system includes a compression assessment procedure that consists of having the user image the vessel of interest on a transverse plane with the contour detection enabled, and press and release so that data of the vessel under different compression can be acquired. In order for data (detected contour, location and force readings) to be considered for a compression assessment of a given vessel section: • The seed points must be within a predetermined volume. All data points must be at least within 10 mm of each other. • There must be at least 8 data points. • The difference between the minimum and maximum applied forces must be at least 3 N. Once data has been collected, the estimates of the transverse vessel area are used to calculate two vessel compressibility measures to determine the possibility of DVT within the examined vessel. The first DVT likelihood measure is called the Transverse Area Ratio (TAR), which is the ratio of minimum to maximum transverse vessel area of a vessel segment. A healthy vessel is expected to compress completely, generating a TAR close to 0%, while it is expected that a diseased vessel will not (TAR  100%). The second measure is the slope of the line fit to the normalized calculated  transverse areas as a function of the normalized applied force for a vessel segment. The slope corresponding to a healthy vein would have a value close to -1, while the presence of a thrombus would be indicated by a value close to 0. 22  2.3  State of the Art  2.3.1  Image Processing  There are many image processing techniques for different ultrasound imaging modalities. These methods are suited for extracting appropriate information from each modality, such as anatomical data from B-mode images and physiological data from Doppler and colour images. The information needed for DVT detection can be described as detecting a vessel contour in a B-mode ultrasound image, and tracking the location of that contour over time. Vessel segmentation can be used in order to construct an accurate quantitative measurement of the vessel morphology [70]. This problem is similar to others which involve feature extraction from ultrasound images, such as contour extraction of the prostate, liver, ovaries or fetal heads. Methodologies for all cases have been reported in the literature. A broad review of ultrasound segmentation methods was recently presented in [99], which focuses on B-mode imaging and specific clinical application areas (echocardiography, breast ultrasound, transrectal ultrasound (TRUS), and intravascular ultrasound (IVUS)). Multiple methods that use assumed feature geometry are presented. In [99], segmentation methods were classified based on whether the output was a detected contour or curve, or a grouping of pixels into regions, among other criteria such as dimensionality of the data. In the 2-D domain, multiple contour finding approaches are based on the active contours concept or active shape models (ASM) and their variations [16, 83]. In ASMs the contour location is determined by balancing ‘forces’ obtained from processed image data acting as external forces and counteracting local constraints determined by the active contour model, such as smoothness or curvature, acting as internal forces. The model is driven by these two types of forces and a feature is detected when the addition of forces is minimized. Chen et al [23] made use of a discrete active contour model to segment ultrasound images of breast tumors. A region of interest must be manually selected and the ultrasound image is preprocessed using Gaussian blurring. An automatic thresholding scheme is implemented, followed by opening and closing operations to serve as initialization data for the active contour models. Segmentation of a 3-D dataset is performed offline, with reported processing time of over 1 minute for the complete segmentation of a tumor. In [92], a Discrete Dynamic Contour model is used to determine the lumen contour of the carotid artery by internal forces calculated from geometrical properties of the contour, and by external forces obtained from the image gray level features. The algorithm is initialized by a single seed point, and the image gradient and the local gray level ratio between the inside and outside of the contour are combined to calculate the external forces. Good results are reported, but algorithm robustness depends on the quality of training images used to obtain algorithm parameters. A vessel segmentation method with active contours was presented by Hern´andez-Hoyos et al [70], which also includes a vessel centerline extraction. An expansible skeleton method is used to obtain the vessel centerline from a 3-D dataset, and an ASM is used to extract the vessel contour from the planes perpendicular to the centerline. Disadvantages of this method are that the centerline tracking cannot deal with vessel bifurcations and the procedure is performed offline. The authors 23  do suggest that more reliable vessel detection could be obtained if the contour detection was not limited to the 2-D planes and included some 3-D information. When attempting to reconstruct a 3-D vascular model, the problem can be decomposed into a series of 2-D cross-sectional segmentations, as was done by Wang et al [133]. A 2-D active shape model was developed to which a balloon force was added as an inflating force. Once a 3-D model was constructed, fluid flow calculations for the vessel were performed using the extracted model and a stabilized finite element method was used to solve the equations governing flow. This method has been validated using analytical solutions and in vitro experiment data. One drawback to this type of method is that the 2-D segmentation does not provide control over 3-D smoothness. No computing time was reported. Another deformable model based method uses a modified balloon model [22]. In this case a combination of the gradient and the second order derivative of the image is used as the external force, while an internal ‘pressure’ is used to push the boundary outward. A 3-D arterial model was constructed using the extracted contours. Good results are also reported for carotid artery images. For most cases ASM’s are quite accurate for detecting the desired features, but even so several disadvantages arise. In many cases in order for the algorithms to converge careful manual initialization is necessary, or as in some cases presented here, an additional processing step for initialization is included. Moreover, as these are optimization procedures, the convergence time may be variable and not the best option for a real-time application. A balance between good boundary detection and good feature tracking over several image frames is also difficult to achieve. Additionally, many active contour methods are relatively complex, and posing another difficult for real-time implementation. Many authors have addressed the contour detection problem in IVUS images [99]. While IVUS depicts vessels, several key differences exist between IVUS images and the deep vein images obtained during a compression examination. The geometry of IVUS images is described in polar coordinates, while linear arrays are typically used for CUS exams. Also, the range of isonation frequencies for IVUS is 30-40 MHz, while CUS exams use arrays at 4-10 MHz, resulting in clear differences in image scale, resolution, and depth. There also exist key differences between image artifacts, such as the void at the centre of IVUS images where the transducer is located, that differentiate IVUS images from those obtained with linear probes. An example of a IVUS image is shown in Figure 2.6 alongside an image from a linear array. Methods used for segmenting vessels in IVUS cannot therefore be directly applied to segmentation of the deep vessels of the lower limb. Some feature detection algorithms include some previous knowledge of the feature’s characteristics. While ASM’s are model based and use a general description of the feature, other methods can include statistical information about the feature as an extension of prior information. An example of the use of this type of method was recently presented by Shen et al [120], where a statistical shape model is used to extract prostate boundaries from ultrasound images. A Gabor filter bank characterizes the prostate boundary in multiple scales and multiple orientations, and then the prostate features are reconstructed and incorporated into a statistical prostate model. This model is used to initialize a deformable shape model. The ultrasound image data is preprocessed 24  (a) IVUS image  (b) Linear US array  Figure 2.6: Example of an IVUS image of a vessel (a) and a vessel imaged using a linear US array (b). Key differences include the coordinate system, resolution, and image artifacts (a small black circle at the centre of (a) where the catheter is). using a Gaussian filter, included in the Gabor filters. This method works offline and segmentation takes almost 130 seconds per image, making it inadequate for real-time applications. A feature detection method was presented by Abolmaesumi et al [1, 3], where a Star-Kalman algorithm was developed in order to detect the contour of a carotid artery from an ultrasound image. The algorithm assumes that the underlying model of the cross section of the vessel is a circle, and detects the most probable location of the vessel lumen based on a probabilistic edge detection function applied to image intensity data. This method has low processing requirements and good performance results. This method was extended using an ellipse model [60] and was used in the system outlined in Section 2.2.3. Doppler or colour flow imaging modalities typically lead to different types of segmentation problems. Even though these imaging modalities obtain primarily physiological data, some anatomical data can be observed in these images, such as the vessel outline for healthy vessels which fill with colour. In 1995, Bell et al [35] developed a technique to segment colour information to estimate the original mean Doppler frequency shift data from which the image was created. An analysis of the colour bars displayed on the ultrasound scanner was performed assuming that flow velocity was proportional to the distance along each colour bar from the centre, and a look-up table was generated to map colours to velocity. It is possible to isolate the flow component simply by isolating an appropriate colour component and subtracting it from the grayscale images. Processing speed for their method at the time was 8 seconds per image, with an expected processing time of about 3 seconds on a 486 processor. A method for 3-D rendering of vasculature of breast tumors, prostate, and liver was developed by Downey et al [38]. 2-D power Doppler video images were digitized while the ultrasound 25  transducer was mechanically moved over a volume of interest. The resulting 3-D image data could then be rendered in a conventional manner. This procedure was performed offline. Similar 3-D reconstruction of vasculature from power Doppler data was reported in [138]. However, data display depended on rendering to provide the structure of the vessels and no geometric models were constructed. Also addressed in [99] is the use of filters in order to improve segmentation. Other examples include filtering of ultrasound images by low pass filters [51] and median filters [60]. When tracking features in ultrasound images over several frames, template matching is a common procedure; the feature to be detected is described by a mask and a correlation procedure is performed to determine its location. [52] presents a comparison of different correlation methods to track speckle motion using Sum of Absolute Differences (SAD), normalized and non-normalized correlation, resulting in similar performance characteristics. [54] also used SAD to track ultrasound speckle patterns to determine blood flow velocity and angle in carotid arteries, and state that ‘SAD methods proved to be the most efficient measure with virtually identical performance to other explored methods.’ Drawbacks to these types of methods include high processing requirements. Once a feature has been segmented, it is desirable to know its accuracy. This can be difficult because of the lack of a definitive gold standard, the difficulty in defining a metric, the lack of standardized protocols and tedious and time consuming data collection. Chalana et al [21] proposed a methodology for the evaluation and comparison of boundary detection algorithms for medical image segmentation. The derived statistical methods enable users to find whether computer generated boundaries agree with observers hand-outlined boundaries as much as the different observers agree with each other, giving a measure of accuracy to the detected features. The limitation of this methodology is that it does not take into account any possible bias of the individual observers.  2.3.2  Location and Force Sensing  Even though many image segmentation methods function in 2-D space, human anatomy exists in 3-D and eventually some type of integration of the acquired information is necessary. Many researchers have identified this and have proposed ways to obtain location information to complement the 2-D image data. Nerney Welch et al [97] reports that an optically tracked freehand 3-D ultrasound system that generates volume rendered images in near real-time has been implemented. A system was developed that allows real-time updates to scanned volume data, which includes an optical tracking system for location data, a video frame grabber and a high performance computer. A 3-D volume reconstruction system aimed at the qualitative evaluation of fetal-placental vasculature has been suggested by Welsh et al [138]. The authors evaluated two 3-D ultrasound reconstruction setups, the first comprising a linear test rig with a mounted tracking device to hold an ultrasound probe, and the second comprising careful manual sweeps for a fully freehand (unsensorized) 3-D ultrasound system. A large error in the plane of image acquisition was found in the freehand setup, enough so that the system would be unreliable for accurate volume measurement, 26  indicating that sensor measurements are necessary. Another example of an application where the ultrasound probe location is tracked is a surgical guidance system developed by Comeau et al [25], where pre-operative image information is combined with intraoperative US to account for tissue deformation. A 6 DOF tracking device is used, either a Surgicom articulated arm or a Polaris freehand optical tracking system. The advantage of the mechanical arm is that the probe is tracked under all circumstances, although its bulk makes it somewhat intrusive in the operating room. The optical Polaris tracker is light and much less intrusive, but its major disadvantage is that it must always remain within the line of sight of the tracking cameras. Many authors have reported successfully using electromagnetic sensors [46] such as in 3-D vascular ultrasound of the carotid artery [47], and reported errors for these types of systems have been low (close to or less than 1% of the dimension of interest)[73]. Additionally, [88] reports that the precision of an Ascension Technology electromagnetic sensor attached to 140 g of aluminum did not change for a low measurement rate (26.5 Hz). Various setups for measuring applied force have been used, such as using an activated mechanical linkage and calculating and controlling the force applied by the linkage from its inverse kinematics. Others have used mechanical linkages with a force sensor directly measuring the force exerted by the end effector, such as is the case with [95, 143].  2.3.3  3-D Vessel Models  Once the extracted information such as vessel contours and the location of these contours is available, a 3-D representation can be constructed. Several authors [38, 104, 138] have generated 3-D representations of vessel data by segmenting flow data, and creating a 3-D volume which is then rendered. These methods can display the morphology of the vessels by appropriately setting rendering parameters, but blood flow and the structure of the vessel is still interpreted by a human observer. There are several 3-D vessel reconstruction applications related to IVUS. In some cases, vessels are reconstructed by simply ‘stacking’ obtained images or segmented contours to create a volume, which gives an inaccurate representation of vessel morphology. By obtaining the location of ultrasound image planes using biplane angiography, Wahle et al [131] reconstructed geometrically correct vessels, overcoming this problem. A similar attempt was presented by Pellot et al [100] using X-ray projections to localize the ultrasound image plane.  2.3.4  Characterization of Physical Properties  The characterization of physical properties of tissue can be performed by a medical imaging technique known as elastography, where measurements of applied force and of resulting displacements are used to quantify tissue properties. It has been suggested that the characterization of the mechanical properties of DVT could improve diagnosis and suggest appropriate treatment [130]. Rubin et al [111] presented their results on relative characterization of clot stiffness, as no force measurements were acquired. Elastography was used to characterize the strain within clots in human subjects, in one subacute case (25 days) and the other a chronic case (more than 3 years). 27  Results indicate the chronic clot was homogeneous and that the strain in the clot was at least 10 times smaller than that in the vessel wall. The subacute case was much more heterogeneous, and on average the strain in the clot was 3 to 4 times larger than the vessel wall. The characterization of clot compressibility for DVT detection using ultrasound elasticity imaging has been reported by Emelianov et al [42], where ultrasound image speckle tracking and strain gauge information is used to determine the age of a thrombus in laboratory rats. The imaged surface was deformed and frame-to-frame motion was estimated using a short time correlation method that employs 2-D correlation-based phase-sensitive speckle tracking, which incorporates the advantages of tracking large displacements from correlation-based techniques with the precision of phase-sensitive techniques. Varying compressibility of clots over a 9 day period have been successfully identified, with clots becoming stiffer as they mature. Young clots are softer than the vessel walls surrounding them, while older clots are stiffer. Much work has been done on vessel characterization from IVUS data, but not specifically for DVT detection. In de Korte et al [36] hard and soft plaques in tissue mimicking phantoms were identified from strain images, independently of the echogenicity contrast between the regions. Wan et al [132] present a system for imaging artery elasticity using an optical flow method, solved using a genetic algorithm. The nature of the iterative minimization for this method still takes too long for consideration in a real-time system. Doyley et al [39] developed a 1-D elasticity imaging technique using global motion compensated cross–correlation to create palpograms from strain profiles. The authors report that the procedure is simple enough to perform in real-time, but experiments were performed offline. In [91], strain imaging was used to measure tissue deformation in response to pressure variations to determine arterial compliance. Precise measurements of changes in area and pressure are needed in order to estimate the cross-sectional compliance of the vessel being examined. Based on their preliminary results the authors suggest that measurements of vessel deformation because of pulsatile flow are a promising way to measure vascular compliance. Until recently, strain imaging techniques had been too time-consuming to be implemented in real-time. A real-time elastography system [142] has been developed that allows the generation of strain images at up to 30 Hz. Moreover, it uses the Sonix (Ultrasonix Medical Corp., Burnaby, Canada) platform and software, allowing easy access.  2.4  Usability  In his book Usability Engineering[98], author J. Nielsen arrives at a definition of usability by breaking down the characteristics of a system: “System acceptability refers to whether a system is good enough to satisfy all the needs and requirements of the users [...] This can be broken down into social acceptability and practical acceptability. Practical acceptability can be broken down into categories such as cost, support, reliability, compatibility, etc., as well as usefulness. Usefulness refers to whether the system can be used to achieve some goal. In turn, usefulness can be broken down into utility and usability. Utility is whether the functionality of the system in principle can do what is needed and usability refers to how well users can use system functionality.” Usability 28  is an attribute of an entire package that makes up a product – hardware, software, menus, icons, messages, manual, quick reference, online help and training. As such, usability testing applies to all parts of a product. Benefits of usability include a reduced learning curve and access to increased functionality for users, and enhanced reputation leading to increased sales, and reduced support and training costs for companies [40]. While usability does incur costs (in the design stage, prototype creation, usability evaluations), the benefits should yield savings that are greater, both from before and after a product is released [40]. For medical devices there is an inherent justification for a very strong focus on usability in order to protect patients from harm, especially when the devices serve a life-critical function [140]. In the United States, the FDA expects manufacturers to explicitly define user requirements and then (1) demonstrate the link between the requirements and (2) conduct tests to determine if users are able to perform tasks effectively. When trying to determine a system’s overall usability, a set of usable measures should be used, which is easier than aiming at “a warm, fuzzy feeling of ‘user friendliness’ ” which may vary considerably from user to user. When defining usability, specific and well defined characteristics and goals should be included, such as learnability (how easy a system is to learn), efficiency (how high a level of productivity can be obtained), memorability (how easy it is to remember how to use), number of errors and satisfaction with the system [98]. The criteria for failure and success should also be clearly defined. There is a general consensus that the most important point for making usable systems is to know the user and the task(s) being performed. Developers should visit a customer site, to have a feel as to how the product will be used. Users should be involved at various stages of the design process, and user-centered design can help to ensure that a final design meets customers’ true needs and preferences. In order to help anticipate learning difficulties and to set appropriate limits for complexity, the class of users that will use the system should be identified, such as work experience, educational level, age, or previous computer experience. The users’ environment and social context can also provide valuable insight, e.g. beeping computers may not be appropriate for a closed office setting. Another important point is to know what the user’s goal is, and what information and tools are needed to reach that goal. Users are busy people trying to accomplish tasks and be productive. Products generally get harder to use as their complexity increases, presenting the users with more features and greater operational demands. In addition to conducting usability test to identify ways to reduce complexity for a specific system, strategies for controlling complexity include: bringing the important features to the surface, eliminating extraneous features, indicating operational status and relevant changes, limiting the number of modes, incorporating protective mechanisms, labeling the important features, using large displays, avoiding coding information, and arranging controls logically [140]. Similar strategies for reducing the chance of use error include: guarding critical controls, confirming critical actions, making critical information legible and readable, simplifying and ensuring proper connections, preventing the disabling of life-critical alarms, indicating and limiting the 29  number of modes, not permitting settings to change automatically, and designing in automatic checks [140]. A usability test is not a research study. The purpose of the usability test is to uncover the most serious problems that users are likely to have with the product. In A Practical Guide to Usability Testing [40], the authors state that not quite half of all major usability problems were detected with 3 participants, 80% of problems were detected with between 4 and 5 participants and 90% were detected with 10 participants. In addition, all global problems were detected with 10 participants. This provides a good starting point for determining the number of users required in a usability study. In Designing Usability Into Medical Products [140], the authors present several generalized points of view provided by nurses and caregivers. In general, nurses want devices to lead them through a clear and consistent series of actions to accomplish a task. In addition, other advice includes alarm systems that cannot be disabled, use large displays and controls, make important features and information prominent, forego extraneous features that only get in the way and have ‘smart’ devices that check if settings fall within boundaries of normal use. System automation is welcome, as long as nurses can remain ‘in the loop’ in terms of understanding a patient’s condition, as is having the means to practice using the medical device before using it on a patient.  2.5  Summary  A survey of DVT detection with ultrasound imaging was presented, along with several image processing methodologies from various areas. It was found that DVT detection with ultrasound can be a reliable process, but many factors can alter the outcome. In addition, many tasks involved in the DVT detection procedure have already been attempted or solved, such as contour detection, probe localization in 3-D space, and quantification of physical properties of tissue, but without a complete system yet available. An overview of usability, which emphasizes including the user in the design of a system, was also presented.  30  Chapter 3  Methodology This Chapter presents the description of the components of the vessel characterization system and is divided into five main Sections. The first two Sections, 3.1 and 3.2, deal with vessel contour detection and tracking. Section 3.3 covers the creation of 3-D vessel models based on the results of the contour detection, and Section 3.4 explores the possible measures that can be used to characterize the scanned vessels using various data. Finally, Section 3.5 describes the integration of the described subsystems into a complete system, including aspects such as the user interface, proposed examination protocols and estimated costs of such a system.  3.1  Contour Detection  Segmentation using spatial Kalman filters has been presented in [1, 3]. In these previous methods, it is assumed that the underlying model is a circle. Image intensity is detected along radii distributed uniformly in a star shape centred at a seed point. This data is used to determine edge measurements based on a probabilistic edge detection function in a spatial Kalman filter having the radius, as a function of radius angle, as a state. This algorithm has been shown to perform well in detecting features even with echo drop-outs and shadowing artifacts. An extension of this circular model using Kalman filters and interacting multiple models (IMM) has been presented in [2] where the same circular model is used to define two systems, that differ in the assumed system and process noise covariances for each model. While this segmentation method has been applied to ultrasound and CT images with good results, the effect is still that of balancing measurements with model prediction, using only one model. This circular model may not be adequate for all cases, including veins under varying degrees of compression or other non-circular contours. It is possible that combining models that differ in structure as well as in assumed noise distributions may yield better results. For the work presented here, an elliptical model was chosen as an approximation of vessels under varying compression, with a range of aspect ratios and orientations, as would be expected of vessel cross-sections during a CUS exam for DVT. It was selected because a few parameters could describe a smooth curve3 and could be used to quickly compute the approximate enclosed area. This approximation will be accurate in many cases, and have significant errors in others. It is not crucial to use an ellipse as the model for the contour detection, since the model is used to provide a good search region for the edge detection, and not to parameterize the edge. In addition to the 2–D contour detection models presented here, the feasibility of implementing 3  And continuously differentiable.  31  a 3–D parameter estimation algorithm using the same approach and an ellipsoid model was also explored. This is described in detail in Appendix B.1.2.  3.1.1  Common Model Features  In general, the contour detection algorithms function as follows. A seed point (xc , yc ) is selected inside the desired contour on an image. N angularly equispaced radii are then projected from the seed point to a maximum search distance rmax , which is larger than the feature to be detected. The radii are sequentially scanned for all θk , where θk = 2πk/N and k = 1 to N . Each radius is filtered by a median filter to improve the edge detection An estimate of the corresponding state vector is then obtained at each k based on the state-space description for a specific model. The estimation procedure is performed by consecutively traversing the contour several times, using the same seed point. The measurement residual zk+1 (described by (A.3) in Appendix A.1) for all models is obtained using a one-dimensional edge detector, which processes the brightness values along the radius rk emanating at angle θk . M number of candidate points ρi are selected based on the results of an edge detection function, and a probability distribution function describing the actual location of the edge is constructed from these points for each radius, using equations (7) to (9) from [1]. In order to insure smoothness the detected contour rˆk and the estimated ellipse are compared to each other. The root mean squared (rms) radial distance between boundaries [21] is used as an error measure, and is computed by measuring the distance between the generated points rˆk and the corresponding points on the generated ellipse. A data fit is deemed invalid if this error is larger than a predetermined threshold, defined by error threshold = max erRM S =  √  1.5 · rmax .  (3.1)  This scaling by rmax is necessary because of the changes in the estimated ellipse parameters and contour points due to initial conditions, and allows a global threshold to be determined for varying search distances. If a data fit is invalid, the contour detection is repeated using a smaller rmax until a predetermined absolute minimum search area is reached. If data is still invalid, the contour detection has failed.  3.1.2  3–Parameter Ellipse Models  An elliptical model has been proposed [60, 63, 64] to describe the contour of a vein and approximate the cross-sectional vessel area. This elliptical model is used as the basis for the development of different contour detection algorithms. An uncompressed vein is described by an ellipse with eccentricity e =  1 − b2 /a2 ≈ 0 which resembles a circle, while a heavily compressed vein can be  described using an ellipse with an eccentricity value e ∼ 1. An ellipse radius r can be described in polar coordinates by r=  ab b2 cos(θ − φ)2 + a2 sin(θ − φ)2  (3.2)  32  Figure 3.1: A contour in an ultrasound image can be approximated by using an elliptical model with parameters a, b and φ. where θ is the radius angle, φ is the angle of the semi–major axis a with respect to the image coordinate axis, and b is the semi–minor axis, as illustrated in Figure 3.1. The ellipse area can be directly calculated by a b π. If the parameters a, b and φ are known we can describe the radius length r of an ellipse centered at a point as the output of the function (3.2) using the radius angle θ and the parameters values. However, if the ellipse parameters are not known, they must be determined. These parameters, and subsequently the desired contour, can be estimated using the framework of nonlinear state estimation of an extended Kalman filter [13, 24]. By treating the discretized parameter values, denoted by Θk = [ak , bk , φk ] , as random constant vectors, such as Θk+1 = Θk + ς k  (3.3)  where ς k is an artificial process noise assumed to be zero mean and white with variance Tk , and including them in our system, the extended Kalman filtering procedure can be applied to estimate a state vector which contains Θk as its components. Indeed, by including the discretized parameter values in our state, we are making our system non-linear. Details of this procedure are outlined in Appendix A.1.1. Various approaches of incorporating the ellipse model were explored, including using the derivative of the ellipse model in the system equation (described in detail in Appendix B.1.1). The approaches implemented in the current system are presented here. Output Method This approach is referred to as the Output model because the ellipse equation (3.2) is included as the output equation. For this system, we define a state vector xk = [ak bk φk ]  comprising the 33  discretized parameters, and write the system describing the dynamics of the parameters as xk+1 = A xk + ξ k  (3.4)  vk = C(xk ) + ηk where A is a 3 × 3 identity matrix, and C(xk ) =  ak bk 2  bk cos(θk − φk  )2  = rk 2  + ak sin(θk − φk  (3.5)  )2  which describes the output rk , and the edge location as a function of the ellipse parameters. Additionally, vk describes the measurement vector, and ξ k and ηk are system and measurement zero mean noise sequences with known covariances Q and R, respectively. As detailed in Appendix A.1, the ellipse parameters can be estimated for all θk using the extended Kalman filtering framework, using (A.2) or x ˆk+1 = x ˆk+1|k + Gk+1 z k+1  (3.6)  where the measurement residual z k+1 is calculated as described in Section 3.1.1. The detected edge rk is the output (3.5) evaluated using the estimated parameters at each k. Since we have a linear system equation for this model, the state prediction covariance Pk+1|k can be calculated directly using (A.12). However, we must obtain the Jacobian of C(xk ), or Hk+1 , in order to calculate the Kalman gain Gk . For this particular case,  Hk+1     =    ∂ C (xk ) ∂a ∂ C ( xk ) ∂b ∂ C ( xk ) ∂φ   a =a ˆk+1|k    . b =ˆbk+1|k    φ =φˆk+1|k  (3.7)  Output Model without Ellipse Rotation Angle There may be cases where it is valid to assume that the contour can be approximated by an ellipse without any rotation, that is φ = 0, and we can therefore simplify the system. An alternative state vector xk = [ak bk ]  is defined for the system in (3.4) which does not  include the ellipse angle φk . The resulting output is simplified to C(xk ) =  ak bk  (3.8)  bk 2 cos(θk )2 + ak 2 sin(θk )2  and the corresponding Jacobian is reduced to   Hk+1 =   ∂ C (xk ) ∂a ∂ C (xk ) ∂b   a =a ˆk+1|k   ˆ b =bk+1|k  (3.9)  34  Figure 3.2: Oscillating estimates of a (‘+’) and b (‘– –’), and true parameter value (‘–’, 70 for both a and b) over 4 cycles. both of which are independent of the ellipse angle φ. As this angle is not included in the state, only estimates of the ellipse semi-major and semi-minor axes are obtained. The corresponding edge rk is still calculated as described above, now using (3.8) with φk set to 0. Alternately, another value for φ can be used, if known.  3.1.3  5–Parameter Ellipse Models  A drawback of the 3–parameter ellipse model is the assumption that the ellipse centre is known, when this may not be the case. One result of this flawed assumption are incorrect or oscillating estimates of a, b, as shown in Figure 3.2. Determining the values of the ellipse centre is therefore important. An initial attempt to determine these parameters was based on the approach used for the 3– parameter model, where, given an initial seed point, N angularly equispaced radii are projected from this seed point and a model is used to describe the length of these radii. Assuming the seed point (xc , yc ) is different from the ellipse centre (xe , ye ), the ellipse model can be used to approximate the contour as shown in Figure 3.3(a), where a, b and φ are as before, θ is the angle for the radius r, and ρ is the distance from the seed point to the ellipse edge at angles αk = 1 to 2π.  35  (a) 5–Parameter Ellipse Model  (b) 5–Parameter Ellipse Model  Figure 3.3: Ellipse with parameters a, b, φ, xe and ye . Distance from seed point (xc , yc ) to contour described by ρ and angle α (a) or θ (b). In (a), if ellipse is centered at (xc , yc ), then r = ρ, and θ = α. The contour can then be described by ρk =  a2 b2 b2 cos(θ−φ)2 +a2 sin(θ−φ)2  + (xc − xe ) cos αk + . . .  (yc − ye ) sin αk + (xc − xe )2 + (yc − ye )2  1/2  + ...  (3.10)  (xc − xe ) cos αk + (yc − ye ) sin αk and assuming that we know the values of the ellipse parameters a, b, φ, xe and ye we can evaluate (3.10) for all k. However, an additional unknown is introduced as we also need the value of θ for each ρ. In the 3–parameter ellipse model, θ and αk as represented in Figure 3.3(a) were equal, but because we are now estimating the ellipse centre as well, θ is unknown. Attempts to include θ in the state to obtain an estimate failed to yield decent results. An alternate approach was therefore sought. By describing the intersection of an ellipse and a line, a solution was found. With respect to Figure 3.3(b), any point x on the ellipse centered at xe , with semi–major axis a at an angle φ with respect to the image axis and semi–minor axis b, satisfies the following xE P R P P P R xE = 1 ,  (3.11)  where xE = [x − xe ] =  x − xe y − ye  ,  (3.12)  36  and PR =  cos φ  sin φ  − sin φ  cos φ  ,  PP =  1/a2  0  0  1/b2  .  (3.13)  If the origin of the coordinate system xc is inside the ellipse, the line with angle θ that passes through the origin also intersects points xP1 , at a distance ρ, and xP2 as shown in Figure 3.3(b). We also know that xP1 = ρ v = ρ [cos θ sin θ] .  (3.14)  By setting x = xP1 in (3.11) and defining P = P R P P P R we arrive at ρ2 v P v − 2 ρ xe P v + xe P xe = 1  (3.15)  and the solution for ρ is obtained using the quadratic formula, simplified to  ρ(a, b, φ, xe , θ)1,2 =  xe P v ±  (xe P v)2 − (v P v) xe P xe − 1 v P v  .  (3.16)  If xc is inside the ellipse we obtain a positive and a negative solution for ρ, and we choose ρ > 0. Eq. (3.16) can be used for contour detection using a radial search and the framework of nonlinear state estimation of an extended Kalman filter (EKF) as was done for the 3–parameter ellipse models in Section 3.1.2. The five ellipse parameters are treated as discretized values and are included in the state vector as before and we define our state as xk = [ak , bk , φk , xek , yek ] and the system xk+1 = A xk + ξ k  (3.17)  vk = Ck (xk ) + ηk  where A is a 5 × 5 identity matrix, Ck (xk ) = ρ(xk , θk ), θk = 2πk/N , vk is the measurement vector, and ξ k and ηk are system and measurement zero mean noise sequences with known covariances Q and R, respectively. The estimate of the state x ˆk+1 is obtained using the EKF equations described in detail in Appendix A.1 for k = 1 to N , and the detected edge is the output ρk from (3.16) evaluated using the 5 estimated parameters in x ˆN and θk . The transverse vessel area can be again ˆ approximated using a ˆ b π. An additional step must be included to determine the positive solution for (3.16), which we will denote as ρk+ , and which will be used to calculate the corresponding output equation Ck+ (xk ) and Jacobian   Hk+1         =        ∂ Ck+ (xk ) ∂a    a =a ˆk+1|k     b =ˆbk+1|k   ∂ Ck+ (xk )  ∂φ φ =φˆk+1|k    ∂ Ck+ (xk )  ∂xe xe =x ˆek+1|k    ∂ Ck+ (xk ) ∂ye ye =yˆek+1|k ∂ Ck+ (xk ) ∂b  (3.18)  37  (a) Direct Sweep  (b) Staggered Sweep  Figure 3.4: Image data displayed in radial coordinates using a small angular step size (b) and a large step size (c). Detected contour dipslayed with ‘o’, and M = 8, N = 64. used for the state estimation. This additional step should also be able to determine whether a real and positive solution exists, since (3.16) will yield imaginary numbers if the seed point is located outside the estimated ellipse. The contour detection algorithm will exit and indicate that an error has occurred if this is the case.  3.1.4  Staggered Radial Scanning  In previous Star algorithms [1, 51, 60], the N search radii are projected from the seed point, with a small constant angular step of 2π/N between radii. Intensity data is used from each radii to determine the edge location, and to estimate the state vector at each angle when using a Kalman filter approach. The estimation procedure is performed by consecutively traversing the contour from 0 to 2π several times, using the same seed point. These multiple traversals are sometimes necessary in order to obtain parameter convergence. By increasing the step size between radii, it is possible to speed up parameter convergence. In addition, it is desirable not to repeat radii, i.e. to have a different θ each time. This is achieved by dividing each full traversal into M equiangular sections, and then dividing each section into another M equiangular sections, for a total of N = M 2 radii. The angle θk = 2πk/N as before, but k is increased by (M + 1). This is equivalent to having a step size of 2π(M + 1)/N . If k > N , then the equivalent index is found. The angle between each of the N radii however remains 2π/N . In this manner, the same data used for a single traversal with small step size is used for the equivalent of M traversals with large step size. An example of the first indices for M = 8 is k = {0, 9, 18, 27, 36, 45, 63, 8, 17, 26, 35, ...}.  (3.19)  Figures 3.4(a) and 3.4(b) show examples of radial data obtained using the original approach and with the larger angular step size, respectively.  38  3.1.5  Angle-Dependant Measurement Covariance  It is well known that tissue boundaries that lie perpendicular to the ultrasound beam will generate stronger echoes. In the context of vessel detection, this means that vessel boundaries will be more visible (brighter) at the top and bottom of the vessel cross section that on the sides4 . Therefore, it is also reasonable to assume that measurements from edge detection methods will have better performance at these same locations. In terms of the contour detection algorithms presented here, this can be defined as a variable noise in the edge measurements. More specifically, assuming the general system described by xk+1 = Ak (xk ) + ξ k vk = Ck (xk ) + η k  (3.20)  where Ak (xk ) and Ck (xk ) are either linear or non-linear system and measurement matrices, xk is the state, vk is the measurement vector, and ξ k and η k are system and measurement noise, all as previously defined for the specific cases in Sections 3.1.2 and 3.1.3. However, where we had previously defined the system and measurement noise covariances as constant matrices Q and R in the previous algorithms, taking into account the varying edge intensities corresponds to assuming a varying measurement noise covariance, or Rk . More specifically, based on the configuration, Rk is dependent on the radial angle θ. Q remains as an angle-independent covariance. A sinusoidal function was chosen to implement the angle-dependent Rk , Rk = Rmax − | 1 − cos θk | ·Rspread  (3.21)  where Rmax is the maximum assumed covariance, which would correspond to the location where there is least certainty in the edge measurements (vessel boundaries parallel to ultrasound beam), and Rmax − Rspread is the minimum assumed covariance, corresponding to the location with most certainty in the edge measurement (vessel boundaries perpendicular to ultrasound beam). The angle is calculated by θk =  2πk π + N 2  (3.22)  where N is the number of radii, k = 1 to N , and π/2 is used as on offset so that the first radii uses measurements with lower covariance. The use of this offset is described in more detail in Section 3.1.7. An example of the resulting angle dependent measurement covariance is presented in Figure 3.5.  3.1.6  Models with Multiple Measurements  The gradient of the image brightness has been the main source of information when estimating the edge location, and has been used as a single measurement for the contour detection algorithms presented here. However, in the context of a system for detecting DVT, other sources of information are available, and these may be included in the contour detection framework for increased reliability. Two such sources were integrated into the contour detection algorithms for the DVT screening 4  Assuming a linear probe is used, and that the probe surface is at the top of the image plane.  39  Figure 3.5: Measurement covariance vs. angle. Minimum covariance values occur at θ = π/2, 3π/2 corresponding to top and bottom of vessel, while maximum values occur at θ = π, 2π corresponding to the vessel sides. system, namely the use of a radial similarity measure from frame to frame and the use of strain or elastography data. Each of these will be described in turn below. In general, the procedure to include additional measurements in the star-Kalman framework is as follows. Assuming that we have a state-space description xk+1 = Ak (xk ) + ξ k  (3.23)  v k = Ck (xk ) + η k  where Ak (xk ) and Ck (xk ) denote the either linear or non-linear system and measurement equations, xk is the state, v k is the measurement vector, and ξ k and η k are system and measurement noise with corresponding known covariances. In the 3– and 5–parameter ellipse models described above in Sections 3.1.2 and 3.1.3, the dimension of the measurement vector v k is 1 × 1. In order to include additional measurements, the dimension of the measurement vector must grow, as must the measurement equation Ck (xk ), the measurement noise vector η k and corresponding covariance R. For example, to add a measurement to the 3–parameter ellipse model from Section 3.1.2 using the output method, the dimension of the measurement vector would be 2 × 1 or v k = [v1k v2k ]  (3.24)  and the resulting output equation would be  q    C(xk ) =     bk 2  ak bk cos(θk −φk )2 +ak 2   sin(θk −φk )2  ak bk q  bk 2        (3.25)  cos(θk −φk )2 +ak 2 sin(θk −φk )2  40  assuming that the measurement is also of the edge location. If the second measurement is not of the edge location, then the appropriate measurement equation should be used. The extended Kalman filter can still be used to estimate the state in the same manner as in Section 3.1.2, with the only other significant difference being the structure of the corresponding Jacobian Hk+1 needed for the linearization of the system   Hk+1         =        ∂ C (xk )(1) ∂a  ∂ C (xk )(1) ∂b  a =a ˆk+1|k  b =ˆbk+1|k  ∂ C (xk )(1) ∂φ  ∂ C (xk )(2) ∂a  ∂ C (xk )(2) ∂b  ∂ C (xk )(2)  φ =φˆk+1|k  ∂φ    a =a ˆk+1|k        b =ˆbk+1|k       ˆ φ =φk+1|k  (3.26)  where C(xk )(i) corresponds the the ith row of C(xk ). The same can be done with the other ellipse models. In addition, other sources of information can be included as measurements for contour detection, and can be added to the model by extending the measurement equation (3.25) and performing the changes described here. While these additional measurements were developed with a DVT screening system in mind, their use in other applications where there is continuous tracking of a vessel is quite possible. Radial Similarity as a Measurement The inclusion of a radial similarity measurement in the contour detection algorithm stems from the fact that the contour detection will be applied to consecutive image frames and the differences between two consecutive frames in a sequence are not expected to be great. Therefore, the pixel brightness values for a given contour point k and the pixels along the k th radii should be quite similar from one frame to another. By calculating a similarity measure between the current and previous frames based on image brightness and the previously detected edge locations, we can propagate information from one frame to the next and ensure contours from these frames are similar. Using the method outlined above, a radial similarity measure corresponding to the edge location was included in the contour detection algorithm. For each radius, an image mask using the pixel brightness from the previous frame is constructed centered at the previously detected contour. This mask is then shifted up and down the current radii by a predetermined search distance, calculating the sum of squared differences between pixel intensities from the mask and the current image data at each location. The location with minimum sum of squared differences, our similarity measure, will be used as the measurement v2k for this radii. Let us assume that at the current frame i we have contour and image data from the previous frame i − 1. Using this radial data r i−1 and image data I i−1 of size rmax × N and in radial coordinates (as illustrated in Figure 3.4(a)), we define a mask mk for each current radii k based 41  (a)  (b)  Figure 3.6: Construction of mask for similarity measure. Given the contour location rki−1 from the previous frame i for the current radius k, a mask is constructed from the pixel values of the previous image using a predetermined width 2Ws + 1. In this example, W = 3. on the location of the previously detected contour as mk = I i−1 [(rki−1 − Ws ) : (rki−1 + Ws ), k]  (3.27)  where rki−1 is the k th contour point from the previous frame and Ws defines the mask window size as 2Ws + 1. This is also illustrated in Figure 3.6. For each radius k, we then calculate n similarity measure values using the sum of squared differences between the mask from the previous frame and the current image data sn = (mk − mI n ) (mk − mI n ) = Σ(mk − mI n )2  (3.28)  where mI n is the radial image data from the current frame and n is defined by the search area window Cs , so n = 0 to 2Cs . More specifically mI n = I i [(qn − Ws ) : (qn + Ws ), k]  (3.29)  qn = rki−1 + n − Cs .  (3.30)  where  The minimum value smin from all sn is found, and the corresponding qmin will be used as the measurement in the Kalman filter, or v2k = qmin  (3.31)  In addition to updating the measurement equation and corresponding Jacobian (as done in (3.25) and (3.26)), the covariance R of this measurement must also be updated. 42  (a)  (b)  (c)  (d)  (e)  (f)  Figure 3.7: Examples of strain images (a), (b), (c) with B-mode data (d), (e), (f) acquired at the same time. A hard inclusion can clearly be seen in the strain image (a) although it does not show up as clearly on the B-mode (d). Also shown are images of the carotid artery. Strain as a Measurement It is possible that two or more different types of modalities may available when imaging a patient. One such modality that may help with contour detection is elastography or strain imaging. By detecting small changes from one ultrasound image frame to the next as the imaged tissue is mechanically excited, local tissue displacement determined from speckle tracking can be used to construct a strain map or strain image of the tissue, by mapping strain (how hard or how soft) to pixel brightness or hue. It has been observed experimentally that vessel walls have strain values that differ from the surrounding tissue. Several examples of strain images are shown in Figure 3.7, alongside corresponding B-mode images. The pixel brightness of these strain maps can be used to generate an additional measurement for the contour detection algorithm. The use of strain data in the star-Kalman algorithm was implemented in a manner parallel to that used for B-mode ultrasound data. Given that strain data exists for a frame i, it is processed radially in the same manner as the image data, and a measurement residual z2,k+1 (assuming the strain data is the second measurement) is obtained using a one-dimensional edge detector along radius rk for all angles θk , as described in detail in Appendix A.1. M number of candidate points ρi are selected based on the results of an edge detection function, and a probability distribution 43  function describing the actual location of the edge is constructed from these points for each radius, using equations (7) to (9) from [1]. Note that these candidate points ρi and resulting distributions will not be the same as those described for the image data, in Section 3.1.1. The measurement equation is expanded to include the additional measurements, as was done above, and the Kalman filtering procedure is performed to estimate the edge locations using all the available data.  3.1.7  Control Parameters  While the basic framework and model of the star-Kalman contour detection algorithm remains the same for all situations, there are several parameters that can be control to adapt to different cases. Most of these parameters have been tuned to deal with vessels when searching for DVT, and these may have different values if the algorithm is applied in a different setting. The user can select the ellipse model used in the star-Kalman algorithm, where the choices are to use the 3–parameter model (Section 3.1.2) or the 5–parameter model (Section 3.1.3). The 3–parameter model is simpler and therefore faster, but the 5–parameter model has the added capability of detecting ellipses with off-centre seed point, such as when seed tracking is off. There are three options regarding the use of median filters on the image data. These filters improve edge detection, but may also cause excessive smoothing. The median window is 5 pixels along each radius, and the three user-selectable options are to use a single radius, 3 radii (the current plus the two adjacent ones) or 5 radii (the current one plus four adjacent radii, two on each side). Processing time increases accordingly. Proper values for the system and noise covariance are crucial for reliable edge detection. These parameters have been tuned for the current application, and represented in a global manner, for consistent results at different ultrasound imaging depths. This global representation is needed as contour detection is done using pixels as units, and may change with different imaging depths as the relationship between pixels and mm changes. The values used are presented in detail for the different experiments in Chapter 4. The use of an angle dependent measurement covariance may be toggled on or off by the user. Also, the use of the similarity measure and the strain data described in Section 3.1.6 may also be controlled by the user. Finally, the output of the contour detection algorithm can be set to use one out of three types: a) the raw detected contour, b) the ellipse constructed from the estimated parameters, or c) the average between the detected contour and the ellipse from the estimated parameters. The output is then be used, for example, for model construction and compression assessment in the DVT screening system. Error Threshold for Contour Validation The threshold defined by (3.1) in Section 3.1.1 was determined experimentally in [60], where it was first noted that erRM S /rmax decreased as the search area decreased, as long as the feature was completely contained within the search area. This scaling is necessary because of the changes in the estimated ellipse parameters and contour points due to initial conditions, and allows a global 44  Figure 3.8: Example of vessels in an ultrasound image. Notice how the top and bottom boundaries of the vessels are brighter and more clearly defined than the lateral boundaries. Also included is the orientation of the coordinate axes. threshold to be determined for varying search distances. Through additional experimentation it √ was found that results improved when using a threshold proportional to rmax . Larger threshold values allow better contour detection for less elliptically shaped features, but may also increase the mean error in the segmented contour. If a data fit has been deemed invalid, rmax is decreased by 5-7% of the initial value of rmax and the contour detection is repeated. If rmax reaches a predetermined lower bound (10-13% of initial rmax ), contour detection has failed. Initial Radius Angle The selection of the initial radius angle is important since it can have an impact on the final estimated contour. The measurement taken at the initial radius, in conjunction with the initial values passed to the algorithm, can in effect bias the final parameter estimate especially if the initial measurement has a large error, and this can have an effect on convergence time. The larger the error, the longer the time the algorithm will take to converge to the correct estimate. Since we are limited to the real-time implementation of this contour detection method, we impose an upper limit on the number of iterations to obtain an estimate of the parameters and contour. We therefore prefer faster convergence time, and using a good first measurement can help with this. It is well known that tissue boundaries that are perpendicular to the incident ultrasound beam will generate stronger echoes than if that same boundary was parallel to the ultrasound beam. This can be clearly seen in ultrasound images of vessels, such as Figure 3.8, where the top and bottom regions of the vessel lumen or wall are brighter than the regions at the side. Given the manner in which the edge measurements are generated, it is clear that a stronger image gradient will be obtained along radii that point either up or down from the centre of the vessel than from radii that point either left or right. It is these radii that point up and down that we want to use as our initial measurements. Given the way the image coordinate frame is defined (see Figure 3.8) and the conventional 45  transformation between Cartesian (x, y) and polar (r, θ) coordinates r=  x2 + y 2 , θ = tan  y x  (3.32)  we must add an offset of −π/2 to our θk so that when sweeping θk from k = 0 to N , θ0 will initially point up (negative y)5 . This will help ensure that the initial measurement for the star-Kalman algorithm is a good one.  3.2  Vessel Tracking  When performing contour detection on an image series, the initial seed point (xc0 , yc0 ) employed in the spatial Kalman filter described above must be provided by the user. This seed point is entered manually on the vein in the image by the operator using a mouse or the ultrasound machine pointing device. Without updating the seed point, the vessel motion with respect to the ultrasound transducer causes the spatial Kalman filter to not converge and the vessel centre is lost. Several seed tracking methods were developed to accommodate feature motion, following the approach presented in [3]. This Section presents an overview of different tracking models and model characteristics that were implemented.  3.2.1  Delayed Measurement  In general a seed point location, xi = [xi yi zi ]T , is estimated through successive image frames i by a temporal Kalman filter. For the Kalman filter dynamics, it is assumed that the vein centre moves with constant velocity from frame to frame. If the current seed location is xi , the seed location in the next frame xi+1 can be described by xi and the previous seed location xi−1 using xi+1 = xi + ∆x˙ i = xi + (xi − xi−1 ),  (3.33)  with the seed velocity x˙ i described by x˙ i = (xi − xi−1 )/∆, and ∆ is the discretization step. Typically, Kalman filters use measurements of the state at the current iteration to update the estimate, in this case the seed location. Unfortunately, a reliable measurement of the seed location at the current frame is not available. A reliable measurement could be obtained from a detected contour for example, but since this itself requires a reliable seed point, the result is a catch 22. However, we do have the seed location at the previous frame, or a delayed measurement. The state space description of this system with delayed measurements is written as X i+1 = A X i + ζ i ui = C X i + ϑi 5  (3.34)  An offset of π/2 can also be used.  46  Figure 3.9: Tracking the seed point. An estimate x ˆi of the vessel centre (xci , yci ) along trajectory s is calculated for each image plane Πi using a constant velocity model. The previously obtained vessel centre ci−1 is used as measurement. where A =  2 · I3×3 −I3×3 I3×3  C =  03×3  (3.35)  [ 04×2 I4×4 ]  and In×n is a n by n identity matrix, while 0m×n is an m by n zero matrix. The state is defined by Xi =  xi xi−1  = [ xi yi zi xi−1 yi−1 zi−1 ] ,  (3.36)  ui is the measurement of the seed point locations, and ζ i and ϑi are assumed to be sequences of white, zero-mean, Gaussian process and measurement noise with known covariances, respectively. Note that the resulting dynamic system (3.34) is independent of the discretization step ∆. The temporal Kalman filter trajectory describes the predicted spatial motion s of the vein centre from frame to frame, as shown in Figure 3.9, and the estimated seed point x ˆi can be computed using known Kalman filtering equations [13]. The 4 × 1 measurement vector ui is composed of the centroid ci−1 of the contour detected at time i − 1 as the measurement of xi−1 , while the measurement of zi is set to zero since x ˆi should lie on the plane Πi . The measurement vector is therefore written as ui =  0 ci−1  .  (3.37)  The location sensor measurements are used to compute the required coordinate system transformations so that all coordinates are described in the current image frame Πi . The estimated point x ˆi may or may not lie on the current image frame Πi , as presented in Figure 3.10. This estimated point must be projected onto the current plane to obtain a final seed point for the contour segmentation algorithm. One solution is to project the estimated seed point 47  Figure 3.10: Projection of the estimated seed point x ˆi onto the current image plane Πi . onto the plane of the ultrasound image, using only the x ˆi and yˆi components (i.e. setting zˆi = 0). Alternatively, a line mi can be defined between x ˆi and x ˆi−1 , and the final seed point can be defined as the intersection between line mi and plane Πi , as described in detail on page 49 in Section 3.2.2.  3.2.2  Additional Measurements - Modifications to the Measurement Equation  While using the most reliable measurement is desirable, at times including additional information can help improve the seed point estimation. This Section deals with adding data using the measurement vector ui and modifying the measurement equation, X i+1 = A X i + ζ i ui = C X i + ϑi  (3.38)  by changing the matrix C. Velocity Measurement Just as the seed point velocity was included in the system equation in Section B.2.1, it can also be included in the measurement. A seed velocity measurement x˙ i can be approximated using the location sensor data and x˙ i = (xi − xi−1 )/∆  (3.39)  where xi is a point on the current image plane, xi−1 is the same point on the previous image plane and ∆ is the known time step. Ideally, the point xi should be the seed point, as in many situations (e.g. rotation about the face of the ultrasound probe) different points on the image plane have different velocities. 48  The velocity is included in measurement vector  . ..  ui =   x˙ i .. .       (3.40)  and if the state includes the seed velocity, then the corresponding rows and columns are added to the measurement matrix C, or   ···  ···  ···      C =  · · · I3×3 · · ·  . ··· ··· ···  (3.41)  For example, for the tracking model with seed velocity in the system described in Section B.2.1 the resulting measurement matrix is   0 0 1 01×3 01×3      C =  03×3 I3×3 03×3  . 03×3 03×3 I3×3  (3.42)  Otherwise, if the state does not include the seed velocity but includes both the current and previous seed point, then   ···  ···  ···  ···      C =  · · · 1/∆ · I3×3 −1/∆ · I3×3 · · ·  ··· ··· ··· ···  (3.43)  must be added. The resulting measurement matrix for the system with delayed measurements from Section 3.2.1 is   001  01×3      C =  1/∆ · I3×3 −1/∆ · I3×3  . 03×3 I3×3  (3.44)  3-D Vessel Model One of the necessary steps for scanning a vessel using the DVT screening system is to construct a 3-D vessel model. This procedure is described in detail in Section 3.3. Once a vessel model has been constructed, the information of location of that vessel can be used as an additional measurement for the tracking algorithm. In general, this measurement can be generated as follows. The 3-D vessel model is described as a series of 3-D vertices determined from the contour detection at a given frame. In addition, the centre of gravity of each of these contours is also known. As the user moves the probe along the scanned vessel, these contours are ‘stacked’ and the 3-D model is constructed. At any given time during the scan, the location and size of the ultrasound image is also known. A segment that connects all of the centres of gravity of the contours can be defined. If this segment intersects the current ultrasound plane, then this intersection point can be used as an additional measurement 49  for the tracking algorithm. More specifically, we define a line that passes between two adjacent centres of gravity, denoted by cj−1 and cj . The orientation of the line is given by the normalized vector v between these two points. The parametric representation of this line is [107] x = cj−1 (x) + v (x) t (3.45)  y = cj−1 (y) + v (y) t z = cj−1 (z) + v (z) t  where t is the distance from point cj−1 to another arbitrary point (x, y, z) along our line. The equation for a plane is [107] n (x) (x − x0 ) + n (y) (y − y0 ) + n (z) (3)(z − z0 ) = 0  (3.46)  where n is the normal to plane, (x0 , y0 , z0 ) is a known point on the plane such as the origin, and (x, y, z) is any other point that lies on the plane. By stating that the point (x, y, z) in (3.45) and (3.46) is the same point we can in this manner generate the measurement xmi based on the intersection of the 3-D vessel model and the current ultrasound plane. By introducing (3.45) into (3.46) and solving for t, we obtain t=  n (x) (x0 − cj−1 (x) ) + n (y) (y0 − cj−1 (y) ) + n (z) (z0 − cj−1 (z) ) n (x) v (x) + n (y) v (y) + n (z) v (z)  (3.47)  which describes the distance from cj−1 to a point xmi on the plane, and where (x0 , y0 , z0 ) is the origin of plane. Once we have the value of t, we can use it in (3.45) to find the coordinates (x, y, z) of the intersection point xmi . This seed point measurement can be included in measurement vector   .. .       ui =  x mi   .. .  (3.48)  and the measurement matrix C is modified appropriately by making this a measurement of the current seed point. For example the resulting measurement matrix, assuming that only this measurement obtained from the 3-D model was added, for the tracking model that includes the current and previous seed point with delayed measurements from Section 3.2.1 is   0 0 1 01×3      C =  I3×3 03×3  03×3 I3×3  (3.49)  and for the system with seed velocity in the system described in Section B.2.1 the resulting mea-  50  surement matrix is   0 0 1 01×3 01×3      C =  I3×3 03×3 03×3  . 03×3 03×3 I3×3  (3.50)  Previous Centre of Gravity in Current Frame An alternative to including the seed velocity in the state or as a measurement, the previous seed location can be projected onto the current ultrasound plane to generate a measurement of the current seed location. Given a centre of gravity ci−1 from the contour detected in the previous frame and the known locations of the ultrasound plane in the previous frame i − 1 and current frame i, calculate a measurement by transforming the previous centre of gravity into the current frame, or xpi = T ii−1  ci−1 1  (3.51)  where T ii−1 is the homogeneous transformation between the previous and current ultrasound image planes. The final measurement can be either xpi directly, the (x, y) components of xpi (i.e. the perpendicular projection of xpi onto the current plane), or by defining a line mi between ci−1 and xpi and using the intersection of that line and the current plane as the measurement, as has been described. The preferred method is to use the perpendicular projection of point xpi . The measurement vector and matrix are modified using (3.48) to (3.50) on page 50 in Section 3.2.2. It must be noted that the use of this measurement is limited by certain assumptions. First, it is assumed that either the tracking update rate is high enough or the user’s movements are slow enough that changes in the location of the ultrasound frame are not significant. And second, if the first assumption is not met, then at least the changes in location of the ultrasound plane can be described mostly by translations, not rotations. Otherwise, this seed measurement will be sufficiently unreliable as to interfere with tracking. Compression Data During the final stages of an examination using the DVT screening system, the user must compress the subject’s vessels by pressing down with the ultrasound probe, as described in detail in Section 3.4. This is done to determine whether there is an obstruction within the vessel, using the change in the transverse vessel area as determined by the contour detection algorithm versus the applied force. Typically, as well as being compressed, the examined vessel appears to move up in the ultrasound image as the tissue between the vessel and the probe is also deformed (as can be seen in the image series in Figure 2.3 on page 12). This information can be used to assist in tracking. When a compression exam is being performed the centre of gravity for each contour used and the applied force at that time is saved. Using this information, a seed measurement for the current frame can be generated given the applied force by using a linear data fit of the coordinate data versus the applied force. The relationship between the (x, y) coordinates and the applied force f 51  is modeled by x = a + bf y = c + df  (3.52)  where x and y are vectors containing at least 2 coordinates from the saved compression data, and f is a vector of the same size containing the force data. The parameters a, b, c and d are obtained using a linear least squares fit. Once the parameters are known, the (x, y) coordinates for the current frame can be calculated given the current applied force fi . This measurement xf i can then be included in the tracking. While the current implementation uses only the 2-D coordinates of the saved centres of gravity, other approaches are possible. The 3-D coordinates can be used instead, and a similar relationship between the (x, y, z) coordinates and applied force can be constructed. The final point can then be projected onto the current plane as has been described above. Alternatively, each 3-D coordinate can be first projected onto the current plane, and then the parameters a, b, c and d are determined from the projected data. Once a seed measurement is available for the current frame, the measurement vector and matrix are modified as before using (3.48) to (3.50). Colour Flow Data An additional source of information can be the data generated from the colour flow imaging modality. This modality creates images with colour values that correlate to blood flow within the imaged vessels. Information such as the direction, velocity, phase and power of the flow signal can be obtained. Colour flow data can be used to address the location of the vessels of interest. This information can be used both as an auto-initialization feature (included as future work) and as an additional measurement for the Kalman filter used for seed tracking. A threshold is applied to the colour data so that individual regions of peak flow are determined, and the centres of each of these regions can be used as seed measurements. These regions can also be selected to lie close to the previously detected contour so that a single measurement is used if several regions are found. Additional processing, such as temporal filtering and combining processed flow data with B-mode data can be used to generate a composite measurement for seed location. The main drawback to attempting to use colour flow imaging to obtain a measurement of the vessel location lies in the geometry of the application. We are interested in detecting vessel contours in a transverse vessel plane, and in this case blood flow is generally perpendicular to the image plane. Since the flow information is obtained based on the Doppler principal, this is the worst possible orientation. The probe must be angled so that the image plane intersects the vessel at an acute angle. This is possible in some cases, but not always. Another drawback is that blood flow in veins can be so slow as to not generate a usable signal. This, coupled with signal attenuation due to depth and a sub-optimal imaging angle limits the usefulness for the proposed DVT application. However, arterial blood flow usually generates a good signal and applications where an artery must be tracked could benefit from the use of colour flow data. 52  Y -Coordinate Offset using Applied Force A final modification can be done to each of the 2-D seed measurements (except those generated using compression data6 ) once they have been calculated. This involves using the difference between the applied force from the previous and current image frames to generate an offset for the y image coordinate. This offset of can be simply calculated by of = Kf (fi − fi−1 )  (3.53)  where Kf is a predetermined negative constant that describes an average displacement for a given force, and fi−1 and fi are the forces from the previous and current image frames. Kf must be negative so that an increase in force from one frame to the next will ‘push’ the seed up in the image (in the negative direction along the y axis). The value of Kf will depend on the type of tissue being compressed, but an approximation using the average displacement for known forces for a given tissue will be adequate. Assuming a linear Kf for this application is acceptable because of the range of forces used.  3.2.3  Interacting Multiple Models (IMM)  Several tracking models have been presented, as well as several possible measurements that can be obtained for them. However, implementing a single model that includes all possible variables in the state, with a measurement vector including all the proposed measurements is not feasible. In some cases, the proposed measurements are very specialized (such as when using compression data) and will only be available for a short period of time during tracking. One possible solution is to use an interacting multiple model (IMM) framework. In this approach r models, denoted by Mj , j = 1, ..., r, can defined to describe the dynamics of the seed point during different ‘maneuvers’ and a final estimate can be obtained that uses a weighted sum of the states of each of these models. The weights will depend on assumed system and noise covariances for each model, and a predefined mode transition probabilities. This is a common approach for target tracking applications, and is described in detail in Appendix A.27 . For example, we can use the delayed measurement system with the current and previous seed point in the state described in Section 3.2.1 which only uses the previous centre of gravity as a measurement as our most probable system. We can then define another system as this same system with the addition of the 3-D vessel model measurement from Section 3.2.2. We then calculate the state and probabilities for each model as well as the likelihood function Λji  in order to update the mode probability µji , or the weights for each of the models. If a 3-D  vessel model is available, then we leave µji unaltered. However, if no vessel model exists, we can simply set the contribution from that model to zero, i.e. µ2i = 0 (assuming j = 2 is the tracking 6 The seed point measurements generated using compression data already take the applied force into account. No further modification is therefore necessary. 7 Please note that the index i used in the text to describe the temporal tracking is replaced with the index k in the Appendix.  53  model using the 3-D vessel measurements). The final state estimate and state covariance can then be obtained. There are several advantages to using this approach as opposed to simply switching as needed from a single model with 3-D vessel model measurements to one without. Since the global state and state covariance are updated at each step, there is continuity from one step to the next, allowing for a smooth transition between models and more reliable estimates. Also, for example, large errors in the 3-D vessel measurement that can arise from patient movement may not have such significant effect on tracking since the likelihood function coefficients Λji depend on the error residual between the predicted state and measurement. Large errors lead to smaller coefficients and therefore less of an effect from that model.  3.2.4  2-D vs. 3-D Coordinates in Tracking Model  While tracking has been described using a 3-D reference frame, the models and methods described in this Section can also be implemented using their 2-D counterparts. Location data is still used, but all coordinates are projected onto a common 2-D reference plane. One specific implementation involves using a 3-D coordinate structure while setting all measurements to lie on the z = 0 plane. Estimated values are also constrained in this manner. The advantage of this approach is that the stability of a simple 2-D model is maintained, while still taking advantage of the location measurements to calculate in-plane movement of the tracked point and thereby allow tracking of fast-moving points.  3.2.5  Seed Point Persistence  It is possible that a valid contour will not be detected for a given seed point. In this case, a new seed point measurement is not available for the next frame and tracking will be interrupted. Since this is undesirable, a scheme was developed to track the seed point over multiple frames, even if the contour detection was invalid. Assuming that at frame i the result of the contour detection algorithm from the previous frame is invalid, the measurement corresponding to the seed location is redefined to use ci−2 instead of ˆ i is obtained. For example, the resulting measurement vector for ci−1 and a new state estimate X the system with delayed measurements described in Section 3.2.1 would be ui =  0 ci−2  .  (3.54)  In addition ci−2 will replace ci−1 if other measurements that use the previous centre of gravity are also being calculated. Alternatively, a new centre of gravity ci−1 can be obtained by projecting ci−2 onto the current image plane as has been described above. This procedure can be repeated using the last valid centre of gravity until a valid contour is detected and the seed point is updated, or until a maximum N number of frames have passed. After that, the seed point will be considered lost and a new manual initialization will be required. The addition of this scheme allows the system to track completely compressed veins over a 54  (a) Full mask  (b) Half mask  (c) Fifth mask  Figure 3.11: The location with minimum brightness in an image region is found using the sum of absolutes and all (a), half (b), or one fifth (c) of the pixels in the region. short period of time, which may not be properly detected with the current detection algorithm. Additionally, the tracking is made more reliable, since one invalid contour no longer required user intervention because of a lost seed point.  3.2.6  Seed Correction Using Sub-sampled Sum of Absolutes (SSA)  The estimated seed location from the Kalman filter is adequate for most cases, with two notable exceptions. Since we assume the seed moves with constant velocity, the seed location may drift when imaging at one location (no movement) and cause the seed location to be lost. Also, large errors may occur with abrupt changes, such as when tracking the seed at one location (no movement) and suddenly moving, or when moving the probe and suddenly stopping. To address these issues, a sub-sampled sum of absolutes (SSA) is performed in the area surrounding the seed point (ˆ xi yˆi ) estimated by the temporal Kalman filter described above. The location with minimum brightness (minimum SSA) is found by repeatedly adding the pixel intensity values in an area of a predetermined size and shifting the search area by a predetermined amount.This location of minimum brightness should correspond to the centre of the vessel8 , and is used as (xci , yci ). In a sense, this is equivalent to a sub-sampled sum of absolute differences (SAD) correlation using a mask where all pixels have a small, equal value. The sub-sampling consists of using a fraction of all pixels, such as 1/2 or 1/5 of all pixels within the search area as illustrated in Figure 3.11, which reduces the processing requirements without a noticeable decrease in accuracy. In addition, the resulting SSA can be compared to a threshold based on the known ultrasound gain. By doing so, incorrect contours detected in regions with higher average image intensities (such as tissue) can be eliminated. 8  Vessels appear dark in ultrasound images, surrounded by brighter regions.  55  (a) Raw Image  (b) Image with Linear Weights  (c) Image with Quadratic Weights  (d) Even Weights  (e) Linearly Increasing Weights  (f) Quadraticly Increasing Weights  Figure 3.12: Including weights in the SSA calculation. Examples of image data without weights (a), linearly increasing weights from the centre (b) and quadraticly increasing weights from the centre (c) are shown, where weight values shown in (d)-(f) taken from the diagonal in (a)-(c). In this example, weight values are in the range of 1 to 16 (image brightness is in the range 0 to 255) and image is displayed in false colour. Notice the increase in image intensities in the weighted images, especially near the edges. Weighted SSA Because of the nature of ultrasound images, it is possible that the location of minimum brightness determined by the sub–sampled sum of absolute values (SSA) may not actually lie at the centre of the vessel. Factors such as shadowing, accentuated image attenuation because of TGC settings or imaging deeper vessels can lead to finding a location of minimum brightness outside the desired vessel. This, coupled with the tracking system that assumes a constant seed point velocity, can lead to errors in the seed tracking procedure. To overcome this problem, the pixel values used in the SSA calculation to determine the location of minimum brightness can be weighted so that the centre of the search area, the initial seed point (ˆ xi yˆi ) from the Kalman tracking, has the lowest weight with a value of 1. By weighting the pixels based on the distance to this point, we can keep the information obtained from the tracking and also compensate for the two exceptions noted above. The sub–sampled sum of absolutes can then be calculated as before. Different types of weights can be added to the image data. Two examples, linear and quadratic weights, are shown in Figure 3.12. Given a initial point and a search area, the weight for each pixel can be calculated as the distance from the given pixel to the centre of the search area (linear) or the square of this distance (quadratic). Through experimentation it was found that the quadratic weights perform better than the linear weights, as pixels closer to the initial seed point are given 56  (a) SSA using image with even weights  (b) SSA using linear weighted image data  (c) SSA using quadratic weighted image data  Figure 3.13: Minimum brightness value using the SSA approach on an unweighted image (a) and a weighted images (b), (c). Notice how the locations of minimum brightness (dark blue) in the unweighted data are shifted toward the edges while the areas of minimum brightness are much more localized in the weighted data, and closer to the desired location (centre of the image). more importance than those on the edges, where most of the false minimums occur. The use of a weighted search area is better than other alternatives, such as reducing the search area, mainly because a small search area may still have locations of minimum brightness that do not correspond to the vessel centre in non-optimal images. In addition, in order for a smaller search area to yield good results, the scaling is highly dependent on the size of the feature, which may only be approximately known. This is shown in Figure 3.13.  3.2.7  Tracking Vessel Bifurcations  Vessel bifurcations are an important feature that will be encountered when vessel scans are performed. On their own, neither the contour detection algorithm nor the seed tracking models can deal with vessel bifurcations. The combined system will simply choose one branch based on the established criteria, such as location of minimum brightness determined by SSA, and continue tracking along that branch. A method for identifying when and where a vessel splits is therefore needed. Several approaches for detecting vessel bifurcations have been explored. Methods that yielded less than satisfactory results are outlined in Appendix B.3. The final approach is described below. Bifurcation Detection with Flow Image Measures Colour flow data can be an additional source of information for the vessel bifurcation problem. A human observer can easily identify a region of flow that grows as the bifurcation is approached and splits as the probe passes over. The implemented approach to vessel detection using flow data involves processing the flow velocities obtained from the colour flow image. This data is represented as an image the same size of the B–mode data in the range of 0 to 255, where 128 corresponds to zero velocity, and 0 and 57  (a) Vessel Bifurcation  (b) Colour Flow Data  (c) Thresholded and Labeled Data  Figure 3.14: A vessel bifurcation in B-mode with the detected contour (white circles) from one branch (a), with corresponding colour flow data (b). The thresholded and labeled data clearly shows two regions, one inside the contour and the other outside. The second flow region is used to initialize a new contour. 255 correspond to the maximum velocity (according to the current ultrasound settings) in opposite directions. First, the flow image is thresholded and labeled so that individual regions of non-zero flow are identified, as shown in the example in Figure 3.14. The area of each region is determined, and a region undergoes further processing if the area is larger than a predetermined threshold (e.g. 150 pixels). For each labeled region the centroid is determined. If the centroid lies within a currently detected contour, then no further processing is necessary. If however the centroid lies outside a currently detected contour, then a new contour and seed are initialized, and a vessel bifurcation has been detected. This approach relies on the behavior of the current system, namely that the seed tracking and contour detection will tend to select one branch of the vessel bifurcation. This approach has several other benefits. Since we know the direction of flow (with respect to the ultrasound transducer) we can automatically assign a flow direction to the currently detected contour, and only initialize a new contour if the flow direction of a candidate region is the same. But the opposite can also be done, that is, detecting veins if we are tracking arteries or vice versa. This allows the possibility of building a more complete vascular map during an examination. This functionality however has not yet been fully implemented and is suggested as future work. Another benefit is that aliasing because of incorrect flow settings may not affect the bifurcation decision, since regions can simply be identified as regions of non-zero flow. The drawback is that we may not be able to correctly identify vessel type based on the flow direction. Finally, another benefit of this approach is that it can also be used as an automatic seed initialization procedure. Whether we know the flow direction of the vessel to be tracked or not, the largest flow region can be used to initialize tracking and contour detection. Because of the periodic nature of blood flow, the use of temporal filtering of the colour data proves useful. The current implementation uses the included ‘persist’ function of the ultrasound engine so that flow regions are more clearly defined and remain consistent over time. 58  Figure 3.15: A seed point for artery (A) is tracked and used to calculate an estimated seed point for a vein (B), which is placed outside the previous contour of the artery for a final vein seed point (C).  3.2.8  Vessel Specific Tracking for DVT  Tracking a vein during a DVT examination has a fundamental limitation in that we expect the vein to completely disappear during a compression exam. The artery however will remain visible when the same force is applied, and can be tracked much more easily. A tracking model that takes advantage of this fact and explicitly tracks a vein and an artery, as opposed to simply tracking a vessel in general. In this model only the artery is actively tracked, using one of the single vessel tracking methods presented here, and calculates the location of the vein based on the initial user input. Once the user has indicated to the system where the vein and artery are located with respect to each other, the system tracks the artery and calculates the corresponding vein location using it’s previous position with respect to the artery. Small changes in the veins position with respect to the artery are allowed, such as the distance or angle between the vessels (which may vary based on the applied force), but in general the positions of the vessels with respect to each other remain the same. Given the initial locations of the vein and artery, the artery is tracked using a single vessel model. For each frame, a predicted vein position is calculated based on the previous pose with respect to the artery, adjusted using the change in the applied force from the previous to the current frame, as described in Section 3.2.2. This predicted position is then corrected by searching in the neighborhood for the location of minimum brightness using a SSA approach as described in Section 3.2.6. This estimated position is used as the final seed point for the vein unless it lies inside the previously detected artery, in which case the seed point is modified by placing it outside the artery contour, along the same line that connects the first estimates of the artery and vein seed points (see Figure 3.15). The distance from the artery seed point to the vein seed point is calculated based on the ellipse parameters estimated for the artery contour, using a proportion of the maximum estimated ellipse axis, such as 1.2 max(a, b).  59  This model addresses several issues. By tracking the vessel that is more reliable9 there is a greater chance that the tracked vessels will not be lost. This is especially important in the context of a DVT examination because of the compression examination. Furthermore, if the vein is not detected, reinitialization is much easier if we still know the location of the artery. Another issue is that since we are forcing the vein seed point outside of the artery, we do not falsely identify the artery as the vein once the vein has disappeared, which can be a problem with other single vessel tracking models.  3.2.9  Control Parameters  Tracking was implemented in the DVT screening system so that many options could be turned on or off. Advanced users can select which tracking model is to be used, as well as which measurements should be included (such as 3–D vessel model measurements, the previous point or compression data) and if the seed point should be corrected using SSA or biased SSA. A development version for detecting vessel bifurcations is also available. In terms of tracking parameters, the most important are the assumed system and measurement noise covariances for the models. These covariances are defined in pixels based on the results obtained from the contour detection accuracy (Section 4.1) and then converted to mm, since tracking is done using metric units, based on the current ultrasound imaging depth. The known scaling factors are obtained from the ultrasound machine to convert the values from pixels to mm.  3.3  Model Construction and Display  From the perspective of the user, a main task when scanning with the DVT screening system is to build a 3-D vessel model. The user must locate the vessel of interest and initialize the contour detection and tracking. Then, the user clicks on a button to initialize the model building process, and drags the ultrasound probe along the vessel in order to construct the model. In order for a detected contour to be used in a 3-D vessel model, two criteria must be met, as described in [60]. First, the applied force must be minimal, to correspond to a vessel under no compression. Second, the location of the centre of gravity cm of the contour must be far enough away from the previously accepted contour cm−1 (unless it is the first contour, in which case it is accepted if the force requirement is met). The contours can be added to the model at any location, as long as these two conditions are met. Once the user has completed dragging the probe across the region of interest, the corresponding model button can be toggled off. Once contours have been accepted for use in the 3-D vessel model, the detected contour points are transformed into a 3-D reference frame and used as vertices. Polygons are then created between the contours whose vertices are defined by j+1 {pjk , pjk+1 , pj+1 k+1 , pk }, for 9  j = 1 : N1 − 1 k = 1 : N2  (3.55)  In the sense that it will not possibly disappear during compression.  60  (a) Polygonal Mesh for 3-D Vessel Model  (b) Colour-mapped Vessel Surface  Figure 3.16: A polygonal mesh is generated to describe the surface of the 3-D vessel model (a), and data describing vessel compressibility (or other characteristics) can be mapped to the vessel surface (b). where pjk denotes the k th point on the j th contour, N1 is the maximum number of contours allowed in a model, and N2 is the number of points per contour. We also define pjN2 +1 = pj1 in order to generate the last polygon for a pair of contours. A vessel model is represented in Figure 3.17. Differences in shape between contours are small, therefore k indexed points remain close to each other and large errors, e.g. due to rotation about the vessel axis, are not expected and have not been encountered. An example of the polygonal mesh generated to depict the surface of the vessel is shown in Figure 3.16. Vertices are grouped by contour, and a default value is assigned to each vertex. Afterward during a compression assessment, these values are modified by group so that the compression criteria (described in Section 3.4) can be mapped to the vessel surface as a colour, also shown in Figure 3.16. Changes and improvements to the model construction procedure are detailed below.  3.3.1  Contour Averaging vs. Contour Insertion  In addition to determining whether a new contour cm is far enough from the previous contour cm−1 using the centres of gravity, an additional check is performed to determine if the new contour is far enough from any other contour cm−n . If the new contour cm is far enough in both cases, then it will be directly inserted into the model as before. But if the new contour cm is not far enough from any other contour cm−n (and still far enough from the previous contour cm−1 ) then the contours cm and cm−n will be averaged, and a new contour will not be inserted into the model. This is illustrated in Figure 3.18. This allows a user to generate an initial 3-D vessel model, and then scan over the same region 61  Figure 3.17: Model Mesh. 3-D vessel model is constructed by creating polygons using detected contour points as vertices. of interest and refine the model. It is especially useful if the contour detection did not generate an adequate contour at one location within the region of interest.  3.3.2  Dynamic Contour Sorting  Each time a new contour is added into the model, a sorting direction is determined and all contours are sorted along this direction. In order to determine the sorting direction, m − 1 vectors are defined using the m model contours. The first vector is defined as the vector between the first and the second centre of gravity of the model contours, the second vector is defined as the vector between the second and the third centre of gravity, and so forth. Finally, a vector addition is performed and the largest component of this sum, either the x, y or z will be used as the sorting direction. Once a sorting direction has been determined, all contours are ordered along that sorting direction using the coordinates for the centre of gravity from each contour. This allows scanning of vessels in multiple orientations in the workspace.  3.3.3  Mapping Multiple Data to Model Surface  Instead of assigning a single value to each vertex, an array of values is now assigned. Previously, a single measure (namely the TAR, see Section 3.4) was used to map colour to the 3-D vessel model surface. Now each value in the array corresponds to a different measure: TAR, slope, flow ratio, radial deformation, or other10 . In addition, it is possible not to group the values for each vertex by contour, and each vertex can be assigned and individual value (see Section 3.4.3). On model construction, all values are still set to a default value and updated as the scanning procedure is performed. The user can switch between which data is displayed, or the data will 10  Please see Section 3.4 for more details on these measures.  62  (a) New contour m far enough from other contours  (b) New contour m close to existing contours  (c) Contour Insertion  (d) Contour Averaging  Figure 3.18: A new contour is inserted into a 3-D vessel model if it is far enough from all previous contours (a),(c), but is averaged with an existing contour if it is close enough to that contour (b),(d). switch automatically depending on the procedure being performed. Examples of the different types of data mapped the the vessel model surface are shown in Figure 3.19.  3.3.4  Branching Models  The model building procedure has been extended to allow the construction of vessel models with bifurcations. Two approaches have been developed for building these types of models that depend on the number of vessels being tracked during the model building and are described in detail below. However, a common procedure is used to generate a bifurcating 3-D vessel model once contours have been determined to belong to different branches of the vessel. The first step is to begin with the normal model construction procedure, so that a model for a single segment is built. Then, a contour must be identified on that vessel that will serve as the anchor for the next vessel branch. This anchor will be called the base contour. Once this base contour has been identified and the system is set to add a new branch, the model construction procedure can continue as before, but now the newly added contours will form a second segment, a segment which also includes the previously identified base contour. The model construction procedure is as described above, where the contour points are described 63  (a) Single Colour per Contour  (b) Multiple Colours per Contour  (c) Multiple Colours per Contour with alternate colourmapping  Figure 3.19: Examples of data mapped as colour to the surface of a 3-D vessel model. A single colour for each contour mapped to the vessel surface (a), and multiple colours for each contour using different colourmaps (b), (c). in a 3-D frame and used as vertices for polygons. In this case, all contours belonging to the new vessel branch will all have a unique index so that they can be identified as such. The steps in the model building procedure, include contour sorting and mesh generation, will be performed as before, once for each branch. Using single seed point and contour If a single seed point is being tracked, the user first must indicate which contour on the pre-existing model will be used as the base contour. This is done by moving the ultrasound probe so that the ultrasound image plane is close to the desired location of the base contour, as shown in Figure 3.20(a), and selecting that contour using a keyboard or button command. The seed tracking must be enabled and operational. The system then locates the closest contour in the 3-D vessel model to the current seed point, and selects this closest model contour as the base contour. The system will temporarily highlight the selected contour, as shown in Figure 3.20(b). Now, the user can continue the model building process, and the newly acquired contours will be inserted into the 3-D vessel model forming a new branch. The base contour will belong to both the original vessel branch and the new vessel branch. A branching 3-D vessel model is shown in Figure 3.20(c), and a wire frame version showing the connections between the vertices in Figure 3.20(d). This procedure can be repeated for any contour in the 3-D vessel model. Using multiple seed points and contours If a method for the automatic detection of vessel bifurcations is enabled, the system would automatically generate a branching 3-D vessel model. The procedure is similar to that performed when tracking with a single seed. 64  (a) Ultrasound Plane near Desired Base Contour  (b) Base Contour Highlighted on 3-D Vessel Model  (c) 3-D Vessel Model with Bifurcation  (d) 3-D Vessel Model with Bifurcation - Mesh  Figure 3.20: Constructing a branching 3-D vessel model. A 3-D model is generated, and the location of the base contour for the new branch is selected (a),(b). The model building procedure is then continued, with the added contours forming a new vessel branch (c),(d). If during the model building procedure, a vessel bifurcation is detected the procedure to establish a base contour for a new branch would be activated. Using the current seed point, the closest existing model contour would be selected as the base contour. In addition, a new instance of the seed tracking and contour detection algorithm would be initialized and indexed so that it would belong to the new vessel branch. The user can then continue scanning using multiple seeds and contours, where each is used to insert new contours to the appropriate vessel branch. There is an exception for selecting a base contour from an existing 3-D vessel model, and this occurs when multiple seeds and contours have already been initialized and are enabled but a model has not yet been constructed. In this case, when the model building procedure commences, the detected contours from each instance will be used to generate separate vessel branches. These branches may not be connected to each other. An example of this may be when scanning both vein and artery at the same time, as shown in Figure 3.21.  65  (a)  (b)  Figure 3.21: Constructing a 3-D vessel model with unconnected branches. Independent seed tracking and contour instances were used to generate two unconnected vessel branches.  3.3.5  Additional Virtual Objects  In the 3-D display several other models are presented, not only the constructed 3-D vessel model. These are outlined in this Section. Ultrasound Image Plane A representation of the ultrasound image plane, correctly scaled to the current depth and width of the ultrasound data, is included in the 3-D view port. The location and orientation of this plane is updated at the system frame rate using the sensor measurements. The use of this plane has been described in the ongoing Chapter. In a previous system [60] an ultrasound probe model was also included in the 3-D display but because the complexity of the model significantly increased computation time, a simple tube model indicating the top of the ultrasound plane was implemented in the current system instead. The ultrasound plane, with the tube model indicating the top of the plane, can be seen in Figure 3.22. Colour Scale One of the most important additional models is the colour scale or colour bar located one the leftmost side of the display. The colour bar indicates the range of possible colours that the model surface can take, with text indicating ‘good’ and ‘bad’ results. Text above the colour bar indicates which measure is being displayed. In Figure 3.22(a), the default colour bar for the TAR criterion is displayed.  66  (a)  (b)  (c)  Figure 3.22: Additional 3-D models on the 3-D display. (a) Vessel model with mapped compression data and text labels, reset point and colour bar. (b) Vessel model and ultrasound plane shown with skin surface model and reset point. (c) Vessel model with ‘vessel origin’ sphere. Text Labels The user has the option of displaying a text label for each model contour, with the current value of the data that is mapped as colour. The text will change accordingly as the user selects different data to display. Even though the colour mapping will reveal trends, such as an incompressible contour surrounded by compressible sections, it is known that colour by itself should not be used to give quantitative information to the user [98]. A text label associated to each contour will be useful if the exact value is required by the user. These text labels can be seen in Figure 3.22(a). Reset Point Each time the user initializes the seed tracking and contour detection by clicking on the ultrasound image, this initial point is saved as a ‘reset point.’ This means that if tracking and contour 67  detection fail, the user can move the ultrasound image plane back to that same location, and tracking and contour detection will be re-initialized using this coordinate. The tracking and contour detection is reset if the perpendicular distance between the image plane and the point is less than a predetermined threshold. In order to provide a visual aid indicating the location of this reset point, a 3-D model sphere is displayed in the view port once the tracking and contour detection have been initialized. This allows the user to quickly locate the reset point if tracking is lost. This sphere will not move until the user clicks again on the ultrasound image. An example of the red reset point sphere can be seen in Figures 3.22(a) and 3.22(b). Vessel ‘Origin’ During a scanning procedure or the subsequent exploration and review of the vessel model, the original orientation of the 3-D vessel model may be lost to the user. In order to assist with locating the ‘top’ of the vessel, a small 3-D model sphere is attached to the 3-D vessel model on the first model contour, as determined by the sorting algorithm. Given the current system configuration, this should correspond to the most proximal contour in the vessel model when scanning the deep venous system. A white model sphere can be seen in Figure 3.22(c). Skin Surface Another useful cue for orienting the 3-D vessel model is to display the location of the skin surface with respect to the vessel model. For each model contour, three arbitrary points from the top of the image are saved, corresponding to the location of the ultrasound probe surface. These surface points are meshed together in the same order as the model contours, and included in the display as a model of the skin surface. This model can be toggled on or off by the user. An example of the skin model can be seen in Figure 3.22(b). Debug Models During the development of the system, additional 3-D model axes corresponding to the coordinate system origin, the sensor location, and the image frame origin were also included. These models are not used during normal system operation.  3.3.6  Control Parameters  There are several parameters used in the model construction procedure that have been determined on a trial and error basis, based on extensive laboratory and field testing. Some of these values are outlined here. In order for a contour to be accepted as a model contour, it must be a contour of an uncompressed vessel. It has been determined that for this to be satisfied, the force applied using the ultrasound probe should be less than 3.0 N. From the current probe and sensor configuration, this corresponds to the force applied in y image axis (or the z force/torque sensor axis). 68  The minimum distance between accepted contours is 0.8 cm, along the sorting direction as described in Section 3.3.2. If a contour is this distance away from the previous contour, but closer than 0.5 cm to another contour in the model, then the new contour will be averaged with the old contour, as described in Section 3.3.1. Otherwise the contour will be inserted into the model. If tracking and contour detection has been lost, then the ultrasound image plane must be moved to where the reset point is so that tracking can be reinitialized. If the perpendicular distance between the image plane and the reset point is less than 0.4 cm, then tracking and contour detection will be reinitialized.  3.4  Vessel Characterization  The final step when scanning using the DVT screening system is to measure the physical characteristics of the scanned vessel segment, and determine if these fall within a healthy or diseased range. This step can be performed once a 3-D vessel model of the vessel of interest has been completed, or the model construction and the vessel characterization can be alternately performed, starting with a small 3-D vessel model. The vessel property we are most interested in is the vessel compressibility. As outlined in Chapter 2, the loss of compressibility of a thrombus filled vein under gentle probe pressure is the most accurate, simple and useful diagnostic criterion for the diagnosis of DVT [8]. This lack of venous compressibility accurately indicates high probability of anechoic thrombus. By measuring vessel compressibility using the DVT screening system we can therefore attempt to determine the occurrence of DVT within a scanned vessel. Other properties of the vessel that we are interested include the use of blood flow data to determine normal or abnormal blood flow, and local tissue stiffness values.  3.4.1  Compression Assessment  In Chapter 2, Section 2.2.3, an outline of a compression assessment procedure was presented. The vessels of interest are imaged on a transverse plane and the user gently presses down with the ultrasound probe and releases. Detected contours and the measured applied forces are then used to calculate two compressibility criteria that can indicate the possibility of DVT within the vessel. The first DVT likelihood measure is called the Transverse Area Ratio (TAR), and is defined as the ratio of the minimum (Amin ) to the maximum (Amax ) detected transverse vessel area for a specific segment, or T AR = Amin /Amax .  (3.56)  The TAR indicates how much the vessel area decreases under compression, as a percentage of the original transverse vessel area, and is calculated from several images. A large TAR (∼100%) would therefore indicate an incompressible vein segment and the possibility of DVT, while a small TAR (∼0%) would indicate a normal compressible vein. The second DVT likelihood measure is a vessel stiffness, which, given the availability of transverse images of veins, and of applied transducer forces, can be defined as the relative change in 69  vessel area divided by the relative change in applied transducer force. A linear fit is obtained to data points defined by the normalized vessel area (A) at a given normalized compression force (F ). Several such points are obtained for each compression cycle (when the examiner presses and releases the transducer at one location) and normalized to the maximum area and force values for that cycle. The vessel stiffness measure is the slope m of the fitted line A = mF + b where the y-axis intercept b is not used. A value m  (3.57) 0 indicates venous incompressibility and  possible DVT, while a healthy vein would generate an m  −1.  Acquisition Protocol In early versions of the DVT screening system, all the detected contours and applied force measurements generated as the user pressed down and released were included in the compressibility criteria calculations. It was noted, however, that in many cases as the vessel was completely compressed and disappeared the tracking and contour detection would either fail or jump to an adjacent vessel and continue incorrectly tracking that vessel. The calculated compressibility criteria would therefore be incorrect. Each of these contour detection and tracking failures has its reasons. In the case of jumping to an adjacent vessel, the search for the location of minimum brightness eventually pushes the seed tracking toward the larger, darker region in the image as the compressed vein becomes increasingly smaller. In the case of failed contour detection, the problem is more fundamental. The system functions on the assumption that there is a vessel to be detected in the image. If however the vessel is completely compressed and has disappeared, the contour detection will ultimately fail. In order to address these issues of collecting reliable contour data for use in the calculation of the vessel compressibility measures, a change in acquisition protocol was made. Only the data generated on the downward, compressive motion is included in the compressibility criteria calculation, instead of using all the data from the compression and release. It is conceivable that using this approach could lead to errors in the compressibility measures, especially if a vessel is not compressed enough for it to completely collapse. While this may have a larger effect on the TAR criteria, which uses both the maximum and minimum transverse area, the effect on the slope criteria should be minimal. Since the goal of the slope criteria is to fit a line to the data, we will obtain very similar values even if the data points with smaller areas are excluded. Predicting Compression Data If the seed tracking and contour detection is temporarily lost during a compression assessment and data points have already been collected, a preliminary line can be fit to the available data and a new point can be estimated using the currently applied force. There is a maximum number of points that can be estimated during each compression exam so that a strong bias is not introduced into the data. 70  Cumulative Use of Data The calculation of the compressibility criteria was enhanced by including data from previous compression-release cycles during an examination, instead of only using the data collected from the most recent compression and release. Upon release by the user, the system determines if there is already compression data assigned to the just-examined contour. If there is, that data is included in the new calculation of the criteria. Otherwise, the criteria values are calculated using only the data from this most recent compression-release cycle. In both cases, the data points corresponding to the model data (i.e. completely uncompressed contours) is also included in the calculation of the compressibility measures. If the current model data is made up of several contours that have been averaged, each contour/force measurement pair is used. Outlier Removal As the number of data points increase, so will the likelihood of adding incorrect measurements. In order to deal with more obvious incorrect data, an on-the-fly outlier removal process was included for the calculation of the compressibility criteria. The outlier removal procedure is based on the calculation of the slope criteria. When estimating this slope parameter, we also obtain a measure of the goodness-of-fit Q of this slope (and intercept, which is not used) to the area and force data. By comparing this goodness-of-fit measure to a predetermined threshold, we can set up a simple method to eliminate data points that are likely to be outliers. If the goodness of fit is below the predetermined threshold, then the point that is farthest from the fitted line (measured perpendicular to the line) is found and removed from the dataset. This procedure is repeated until the goodness-of-fit is acceptable, or until only two points are left. However, any data points that have been removed will be used again if more data is acquired for this particular model contour.  3.4.2  Use of Blood Flow Data  In addition to vessel compressibility, determining whether blood flow is unobstructed in a vessel can help confirm or rule out the possibility of DVT. One method conventionally used to determine if flow is normal within a vein is called an augmentation exam. The examiner images the vessel of interest on the long axis with colour flow imaging, and either asks the patient to flex their ankle or the examiner gently squeezes the patient’s calf. A large increase in flow is observed for healthy vessels, while a minimal or non existent flow increase points toward a vessel with DVT. In general, when a patient has healthy veins, an augmented flow signal is observed proximal to the location of compression. The increase in flow is substantial close to the location of the compression, and decreases as the location of the flow measurement moves farther proximally from the location of compression. Even so, the increase in flow should still be noticeable. When a patient has a thrombus distal to the location of the measurement but proximal to the location of 71  compression, the response to the augmentation is greatly reduced. When the thrombus is proximal to the location of the measurement, the response is almost eliminated, accompanied by a widening of the vessel lumen, as the blood pools distal to the thrombus. As a special case, if the measurement is taken at the location of the thrombus, an increase in the velocity of the blood flow may be seen as blood passes by the thrombus. A flow ratio calculation has been introduced by Sasaki et al [113] to measure this increase. Peak flow signal with active maximum ankle flexion and that at rest were measured with continuous wave Doppler and used to calculate flow ratios in the femoral veins. Flow ratios for patients with non occlusive DVT were significantly lower than those without DVT. Mean flow ratios for patients positive for DVT were 1.18 (range 1.0 to 1.3), while mean ratios for those negative for DVT were 3.31 (range 1.8 to 4.8). Using the colour flow imaging modality, a flow characterization component was included in our DVT screening system based on the conventional flow augmentation examination. This procedure consists of acquiring flow data when the patient is at rest and while actions to augment blood flow are performed. The increase in flow can be calculated from the peak flow values from the two states. The procedure is as follows. After the examiner has built a 3–D model of the vessel of interest, the examiner places the ultrasound probe so that the vessel is imaged on the longitudinal axis. The location of the vessel can be determined in several manners, such as using the location data from the 3–D model, the seed point or by manual selection (default). The flow data collection is then initiated, and the system obtains a flow measurement while the patient is at rest. After a predetermined time (∼2 seconds) the system indicates to the examiner and the patient that the patient must either flex their ankle or that the examiner must gently squeeze the calf. After a final predetermined time (∼2 seconds), peak flow values for both patient states are determined, and the amount of flow augmentation is determined by calculating the ratio of maximum flow for each state. The result is then mapped to the appropriate 3–D model slice. This flow assessment can be performed once per vessel segment, as all flow ratios for a segment should be the same, or an average value can be calculated from different assessments for each vessel segment. As such, the system does not require the full 3–D model to be ‘coloured’ for a complete examination, as with the compression assessment. Since the location of the examined vessel has been determined during the 3–D model construction and the current location of the ultrasound plane is known, possible improvements to the system include optimal angle calculation for the ultrasound flow measurement, or requirements of flow measurements at several locations if, for example, a vessel with several branches has been constructed.  3.4.3  Local Contour Properties  Another manner in which to extend vessel characterization is to quantify the properties on a very local level. Specifically, the location of each of the detected contour points can be used to integrate very local information about the vessel being scanned, such as the strain of a small region of vessel wall. Using the subsequent locations of the detected contour points, more complex data can be 72  collected. This information could possibly be of use in several applications. For example, calculated strain could be mapped to the 3-D model surface (see below) when scanning for DVT, or sections of the vessel lumen that move more than others could be displayed if scanning a carotid artery, by calculating the maximum displacement of a specific contour point over several frames. Figure 3.19 shows examples of mapping different data to each of the individual detected contour points. While the system already boasts this capability, potential applications (other than DVT screening) should be explored for visualizing data in this manner. Use of Strain Data If strain data is calculated during a scan using the DVT screening system, this data can also be mapped to a 3–D model of a scanned vessel. Strain data can be mapped in several manners, such as using the mean, maximum, or minimum strain over a predetermined number of frames. For each frame, the calculated strain data can be sampled at the coordinate of each detected contour point (or the average of the pixels in the strain data surrounding the contour point, for smoother data). This data collection is repeated for a predetermined number of frames, and the final value (mean, minimum, maximum) can be assigned to the model location that corresponds to that contour point. The final result is a representation of the scanned vessel that contains information not only along the long axis of the vessel (as is obtained with the TAR and slope criteria), but also local information that may vary for each and every point defining the model, and in this way providing a more complete representation.  3.4.4  Control Parameters  Some parameters used during for vessel characterization are outlined here. The difference in the applied force between two ultrasound image frames is the principal factor for determining whether to accept a detected contour and force measurement for use in calculating the compression criteria. This difference must be greater than 0.7 N to accept a new contour. In addition, all contours must be no farther than 1.0 cm away from each other to be considered part of a same vessel region. In order for the calculated compressibility criteria to be valid, the difference between the applied minimum and maximum force must be greater than 5.0 N, and the dataset should consist of at least 4 points. The maximum number of estimated points per dataset is 2. During the outlier removal procedure, if the goodness-of-fit Q is less than 0.4, then the data point with the largest error is found and removed. A goodness-of-fit to the data is good for values greater than 0.1, and typical values found from laboratory tests are close to 0.9.  3.5  System Integration and User Interface  The DVT Screening System has been implemented using Microsoft Visual Studio, in C++. When possible, third party APIs and libraries with specific functions have been used. Additionally, other 73  Figure 3.23: Hardware diagram of the DVT screening system. free and commercially available function libraries have been included in the system to improve its functionality. For display of the 3-D virtual environment, and for the creating of the 3-D vessel model, the Visualization ToolKit (VTK) [74] has been used. The Intel Math Kernel Library (MKL) [77] has also been used throughout to improve the speed of the system, specifically for the contour detection and seed tracking. The Intel Compiler for Windows [76], version 7.0, was also used to improve system performance. Figure 3.23 shows a schematic diagram of the hardware configuration of the DVT screening system. It is composed of a computer platform and a sensorized ultrasound probe. The computer platform includes the ultrasound imaging engine.  3.5.1  Ultrasonix Platform  The initial prototype of the DVT Screening System was implemented on a dual Intel Pentium III processor PC at 1.13 GHz with 2.0 GB of RAM, a 64MB video card and runs on a Microsoft Windows XP operating system. This was the platform for the Ultrasonix Ergosonix R RP system, software version 1.3.1.2. The current version of the DVT screening system is implemented on a Ultrasonix Sonix R PC-based ultrasound machine with research package. It is a dual core Intel Pentium 4 processor PC, running at 3.00 GHz with 1.0 GB of RAM, on a Microsoft Windows XP operating system. The Sonix R includes the ultrasound control interface, as can be seen in Figure 3.24. An estimate of the cost of a prototype DVT screening system, as well as a commercial system is presented in Appendix C. Ultrasonix provides code and libraries as part of the research package software that allow direct access and control to the ultrasound imaging engine and ultrasound data. In addition, the 74  Figure 3.24: Ultrasonix Sonix R PC-based ultrasound with DVT screening system. Sonix R system runs on a Windows R operating system. This has been pivotal in the development of the DVT screening system for several reasons. The most important is the direct access to the ultrasound data, in the form of function calls that allow code and programs to be written and executed directly on the ultrasound machine. The ultrasound data is accessible as a memory location and can be included as a variable, which eliminates the need for video acquisition hardware and greatly increases system speed. In addition, data is available at the original resolution. Other important uses of the research package software include direct access and control of ultrasound parameters, such as gain, depth, frequency, and control over the system, such as selecting predetermined ultrasound probes and preset imaging modalities (and changing these modalities on-the-fly) all through software. This means that the complete Sonix R ultrasound engine and interface can be controlled and accessed through custom software if desired. The current system processes ultrasound image data at the native frame rate (ranging from about 10 Hz to about 40 Hz), and the virtual environment is displayed at a fixed 20 Hz. Contour detection and tracking of a single vessel can be performed at up to 40 Hz. Up to three seeds/contours can be processed for each image frame. Increasing this number however leads to a noticeable decrease in system efficiency.  75  3.5.2  Sensors  Force Sensor The current implementation uses a force/torque sensor (Nano25, ATI Industrial Automation, Inc.) with 6 DOF. The full scale loads of the force-torque sensor are rated as 125 N for the X and Y axes, 500 N for the Z axis, and a torque of 3 N-m for all three axes. The data is acquired by a 16-bit data acquisition card (PCI-6034E, National Instruments Corp.). Data acquisition rate is coupled to the location sensor acquisition rate of 26.5 Hz. A sensorized hand-held ultrasound probe was developed in order to obtain the necessary measurements for the DVT system [60]. Two aluminum shells were constructed to surround a linear 9-4MHz ultrasound probe. The inner shell is fixed to the probe, while the outer shell is connected to the inner shell through the force/torque sensor at the rear as illustrated in Figures 2.4(a) and 2.5 on pages 21, 22. The examiner can grasp and manipulate the ultrasound probe in a regular manner, and all applied forces and torques are acquired. While the configuration of the probe and shells is not required for the DVT system, what is required is a sensor configuration that will allow the system to measure the user-applied force with minimal modification to the ultrasound scan head. Location Sensor The location measurements for the current DVT system are obtained from an electromagnetic sensor (PCIBird, Ascension Technology Corp.), attached to the rear of the outer probe shell through a 14 cm Plexiglas rod. Measurements are obtained using the library files provided by the manufacturer. The sensor and transmitter are connected to an electronics unit, which connects to the computer through a serial port. The static resolution of the sensor with no interference is specified as 0.5 mm for position measurements and 0.1 degree for angular measurements, with an accuracy of 1.8 mm and 0.5 degrees for position and orientation, respectively, which is within the stated requirements. Since the sensor is maintained in rigid relationship to the ultrasound probe, and therefore the ultrasound image plane, the location of the image plane can be calculated using a calibrated homogeneous transformation and the location measurements. The extracted 2-D contours can therefore also be placed in a 3-D reference frame. With respect to possible interference from the aluminum probe shells, Leotta et al [88] reports that the precision of an Ascension Technology Corporation electromagnetic sensor attached to 140 g of aluminum did not change for a low measurement rate (26.5 Hz). The measurement rate for our system was set to 26.5 Hz, which is adequate for our purposes.  3.5.3  User Interface  Many iterations of the user interface were created. Some of these were made with experimentation in mind, and little attention was paid to the usability of these interfaces. This section however will focus on the user interface used in the usability evaluations of the DVT screening system, described in Section 4.4. 76  Figure 3.25: The screening system interface. The DVT screening system is closely tied to the conventional compression ultrasound examination, and in a sense mimics the process by which the US technicians determine the presence or absence of thrombus. Therefore special attention was paid to these procedures performed by users. Many hours were logged sitting in on conventional examinations, and the input and feedback from ultrasound technicians was considered during the ongoing development of the system. Making the system easy to use was set as a priority. Figure 3.25 shows the DVT screening system interface, which is divided into three main regions: • The first, labeled ‘A,’ is the ultrasound display where the ultrasound data is presented. The detected contour is also shown in this area. • The second region is the 3-D display, labeled ‘B,’ where the 3-D vessel model will appear, as well as a 3-D representation of the ultrasound image plane and other actors (see Section 3.3.5). The ultrasound image plane is presented as a dark gray plane, without any image. • The third region is the control interface, where the buttons are located. There is also a status window, labeled ‘C,’ where the current status of the system will be visible, along with instructions or cues to follow during the exam. This version of the interface has been streamlined for use by the target user group. Only the necessary controls have been included, with one button for each of the main actions to be 77  Figure 3.26: The ‘Help with Scan’ dialog. A short version of scanning instructions for use with the DVT screening system. performed during a scan: ‘Find Vessel’ for contour detection and tracking, ‘Make Model’ for the 3-D vessel model construction, ‘Compr’ss’n Exam’ for assessing a vessel segment for compression, and ‘Colour Flow’ for obtaining flow data and calculating a flow ratio for the given vessel segment. Keyboard shortcuts are also implemented for each of these actions. A dedicated button for user assistance (‘Help with Scan’) is also prominently located in the control interface. The user can select this at anytime during an examination and receive instructions on how to perform a scan. The instructions are contained on three screens. The first is a short version of the instructions (see Figure 3.26). The second, accessed through the first using the ‘More’ button contains more detailed instructions (see Figure 3.27), and the last one is a more detailed description of the flow assessment procedure (see Figure 3.28) accessed using the ‘More’ button in the flow instructions. The instructions for each step are clearly identified by the same icons that are used in the main interface. In addition, other cues appear at important steps of the examination: • When the system is started the force sensor must be biased, so a dialog appears indicating to the user that they should hold the probe out facing downward. This eliminates the possibility of the user forgetting to bias the sensor, and reduces the possibility of an incorrect bias11 . 11  The development version of the software was written assuming an expert would use the system. No indication  78  Figure 3.27: The ‘Help’ dialog. A more detailed version of scanning instructions for use with the DVT screening system. • A reminder dialog appears at the end of the scan, instructing the user to record the patient ID generated by the system. • During a flow assessment, the system clearly displays ‘AT REST’ below the ultrasound image and flashes to ‘FLEX’ when the patient is required to flex. This decreases the mental load on the user, as the system is reminding them what actions or instructions should be given to the patient. Other controls, most of which were implemented for the development and testing of the system, have been limited, but are still accessible through menu items. This is a typical manner to keep a was given as to when the force sensor was biased – it was expected that the user would know. This could lead to an incorrect force sensor bias. The current approach of including a dialog reduces this possibility.  79  Figure 3.28: The ‘Flow Help’ dialog. Detailed instructions for flow assessment using the DVT screening system. system usable for novice users, while retaining system functionality [98]. This added functionality includes changing the point of view for the 3-D vessel display, maximizing the 3-D vessel display so that it can be viewed in full-screen, and control of the options of the various subsystems such as tracking and contour detection.  3.5.4  Scanning Overview  In general, a user scans a vessel by pressing the ‘Start’ button12 and then proceeds to locate a vessel on a transverse plane using the ultrasound probe. Once a vessel has been located, the user must press the ‘Find Vessel’ button and set the initial seed point. The contour detection and tracking then commences. The user should then press the ‘Make Model’ button and lightly drag the probe along the length of the vessel. This will make the 3-D vessel model. If the user is pressing down too hard, the 3-D model will not be built. The user can then turn off the ‘Make Model’ button or press the ‘Compr’ss’n Exam’ button to begin the compression assessment of the vessel. A detected contour should be visible to proceed. The user then presses down gently, and the examined vessel should compress (if healthy). Upon releasing, the system will process the data for this location, and map the compression assessment 12  The displayed text on this button changes to ‘Review’ when selected.  80  results to the corresponding model location. The user should then move the probe to a new location along the vessel, and repeat the compression assessment procedure by pressing the ‘Compr’ss’n Exam’ and pressing down and releasing with the ultrasound probe until the complete vessel has been assessed for compression. Once the complete model has been assessed for compression, the user can press the ‘Colour Flow’ button to assess the blood flow in the vessel. The user should identify the location of venous flow on-screen and select it. The system will display the message ‘AT REST’ followed by ‘FLEX’, indicating to the user that they should instruct the patient to perform these actions. The flow assessment can be done once or multiple times, for each vessel model. In addition, the flow assessment can be done before the compression assessment if so desired. If the detected contour is lost during the time that contour detection and tracking is required (model building, compression exam), the user can reinitialize the system simply by moving the probe back to the location of the reset point (in Section 3.3.5, page 67) or by manually re-locating the vessel on the ultrasound image. Finally the user can select the ‘Review’ to complete the examination. This procedure is repeated for each vessel segment.  3.5.5  Proposed Examination Protocol  The examination protocol is based on the conventional CUS examination, as first presented in [31] and otherwise generally accepted. The examiner should scan the patient lying supine with slight external rotation of the leg. The scan should start high on the thigh of the patient, generally near the level of the groin crease. The examiner should locate the deep venous system, and scan these vessels on a transverse plane, assessing the veins for compression by gently applying pressure to the patient’s limb using the ultrasound probe. Starting at the common femoral vein and moving distally, the examiner should do the following: • Locate the common femoral artery (CFA) and the common femoral vein (CFV). To confirm that the CFV has been located, the superficial femoral vein (SFV) and (long saphenous vein) LSV insertions should be seen on a slightly distal scan. • The CFV should be assessed for compression using the process outlined for compression exams. • Move distally and locate insertion of LSV medially. The proximal portion of the LSV should be assessed for compression. • The CFV should be located again using grayscale, and then move distally to locate insertion of the profunda vein (PROF). This lies posterior and lateral to the SFV. The PROF should be followed as far as possible assessing compression as stated above. • Return to the SFV insertion. The SFV should be followed all the way down thigh, assessing for compressibility at proximal, mid and distal thigh.  81  Figure 3.29: Vessel Viewer Software. Clockwise from left, navigation display, 3–D vessel model display, compression criteria display, and ultrasound image with detected contour. The popliteal fossa should be examined with either the patient prone or supine with the knee bent and with external rotation if there is limited mobility. Also the decubitus position may be used or the leg can be elevated and supported off the couch by an assistant. • The popliteal vein (POP) should be located posterior to the popliteal artery, and assessed for normal compression as outlined. The popliteal vein should be followed as far as possible distally, if possible to level of the trifurcation into calf veins. The largest of these should be followed as far as possible as well, assessing for compression along the whole way. • The posterior tibial vessels (PT) are more easily located scanning in the transverse plane from the medial side of the calf or posteromedial aspect of the leg and identifying the artery. • The anterior tibial veins (AT) are best visualized from the anterolateral approach, with the transducer positioned between the tibia and the fibula, and scanning transversely. The AT vessels are found in the superficial aspect of the interoseous membrane. • The deeper peroneal (PER) can be imaged using the same transducer position as the PT 82  veins. If the same position cannot be used, a straight posterior approach may be tried. As well, the peroneal veins can best be seen from the anterolateral approach, as used to visualize the AT veins. The procedure may be repeated for the opposing limb.  3.5.6  Vessel Viewer  Additional software for viewing collected data was also written, with the assistance of Neerav Patel. The objective of this software was to provide a platform for reviewing scans, displaying compressibility measures, and providing a graphical interface to the collected data. A screen shot of this system is presented in Figure 3.29. Features of this software include a summary of examinations, with images used for model construction and compression assessment, display of detected contours and a force vs. area display for presenting the slope compressibility measure. In addition, the user can point and click on the 3–D model of the scanned vessel to select data to display. Individual data points can also be selected and removed from the compressibility criteria calculation for manual outlier removal13 .  13  Implemented prior to online outlier removal described in Section 3.4.1, page 69.  83  Chapter 4  Experiments and Validation This Chapter presents the simulations and experiments conducted for the evaluation and validation of the DVT screening system and its components. The Chapter is divided into five main Sections. Sections 4.1 and 4.2 describe the validation of the vessel contour detection and tracking algorithms presented in Chapter 3. Section 4.3 covers the validation of the measures for vessel characterization and presents the results from the initial pilot study. Section 4.4 describes the usability evaluation of the DVT screening system and finally Section 4.5 presents the results from the final clinical study evaluation the DVT system.  4.1 4.1.1  Contour Detection 3–Parameter Ellipse Model  Convergence Tests The output model was implemented and tested to see how well the ellipse parameters were estimated, and to determine if the resulting estimates converged. In addition, the number of iterations necessary to obtain adequate estimates from the algorithm was also determined. The same experiment was conducted for the derivative model described in Appendix B.1.1. This was done by generating random ellipse parameter values within a known range and substituting these values into the ellipse equation (3.2) with the known θk value in order to generate a measurement vk =  ab b2  cos(θk − φ)2 + a2 sin(θk − φ)2  (4.1)  and calculate the measurement residual z k+1 = v k+1 − Ck (ˆ xk+1|k )  (4.2)  using the output equation evaluated with the predicted state x ˆk+1|k . The ellipse parameters values were chosen as follows. The semi–axes values, a and b, where chosen in the range [10, 100]14 with a uniform distribution. If the values of a and b were similar (the difference between them was less than 10), then b was chosen again. The angle φ was chosen in the range [−π/4, π/4] ([−45◦ , 45◦ ]), also with a uniform distribution. The initial conditions for the output algorithm was [50 50 0] and [50 50 50 0] for the derivative algorithm. Three benchmarks were established, and the number of iterations to reach each benchmark was recorded. The error between the estimated and true parameters were calculated at the end 14  Arbitrary units were used, however the range in values is similar to what is encountered in a clinical setting.  84  Table 4.1: Convergence benchmarks for 3–parameter ellipse models. Shows the average number of iterations required to reach each benchmark. Algorithm n Benchmark 1 Benchmark 2 Benchmark 3 Output Derivative  (no.of iterations)  (no.of iterations)  (no.of iterations)  1890  2.69  3.43  5.11  3  5.00  5.00  5.67  of each iteration. The errors for the ellipse semi–axes were also normalized by the true parameter values. The benchmarks were 1. When the error for the ellipse semi–axes was less than 10% of the true parameter value and the angle φ error was less than 5◦ . 2. When the error for the ellipse semi–axes was less than 5% of the true parameter value and the angle φ error was less than 2◦ . 3. When the error for the ellipse semi–axes was less than 1% of the true parameter value and the angle φ error was less than 1◦ . In addition the mean squared error (MSE), as well as the root-mean squared error (RMS), between the detected ‘contour’ and the true contour were calculated. The algorithms were executed 2000 times with a maximum of 20 iterations each. If the three benchmarks were not reached within the maximum number of iterations, then it was considered that the algorithm did not converge. The average number of iterations for each of the benchmarks for both the output and derivative methods are presented in Table 4.1. The number of times the algorithms converged is also shown (n). It is quite obvious from Table 4.1 that the derivative approach using the 3–parameter ellipse does not converge. Only 3 out of 2000 evaluations (0.15%) met all three benchmarks in less than 20 iterations. Examining the results after 20 iterations also indicates large errors for the ellipse parameters (up to 800%) and contour errors. Also, 20 iterations for each contour is too much for real-time implementation. The output model, on the other hand, shows good results. Only 110 of 2000 (5.5%) cases did not ‘converge.’ As expected, the ellipse parameter values are less than 1% for a and b and less than 1◦ for φ, as can be seen in Figures 4.1(a) and 4.1(b). In many cases, the errors are very close to zero. In addition, low mean squared errors and root mean squared errors are obtained, as illustrated in Figures 4.1(c) and 4.1(d). Given these results, the derivative method was abandoned and efforts were focused on the output approach. Simulated and Patient Image Data - Procedure The contour detection algorithm was evaluated using images in two steps. The first step was to determine the accuracy of the algorithm for detecting ellipse parameters and contours from 85  (a) Ellipse Parameter Errors for a and b  (b) Ellipse Parameter Errors for φ  (c) Mean Squared Error  (d) Root Mean Squared Error  Figure 4.1: Final ellipse parameter (a), (b) and contour (c), (d) errors obtained during convergence tests for the 3–parameter ellipse model using the output approach. elliptical features with known characteristics, and the second step was to determine the accuracy with which the algorithm detected the actual feature contour determined by expert tracing. For the first step, ultrasound images were simulated using Field II c [80], with elliptical features with known parameters. An example is shown in Figure 4.2. The Field II package provides an excellent framework for simulating ultrasound images using linear acoustics [99], and has been used for various applications [18, 69, 86]. Nine different sets of ellipse parameters (eccentricity e = {0.0, 0.745, 0.866, 0.916, 0.943, 0.958, 0.968, 0.975, 0.979} and φ = 0◦ for all cases) were generated using 3.5 MHz as centre frequency, 100 MHz sampling frequency, 192 elements (64 active elements) and 50 lines per image. The speed of sound was set to 1540 m/s, and data was compressed to show 60 dB of dynamic range. 80,000 scatterers were randomly generated for tissue, and their amplitudes were randomly set with normal distribution and maximum amplitude 1 as per typical Field II simulations. Amplitude was set to 0 inside the defined ellipses. In total, 18 simulated images were used (320×440 pixels, 0.091 mm/pix). Nine images as described above, and 86  Figure 4.2: Examples of contour and parameter estimation. Simulated data used to test contour detection, with expert tracing (a) and detected contour (b) with detected contours (white circles). The estimated semi-major axis a and semi-minor axis b are presented as a solid and dashed line, respectively, oriented according to the estimated φ. an additional 9 images with a 2 mm border, where the amplitude of the scatterers had a normal distribution and maximum amplitude 3. The effect of noise on the accuracy of the algorithm was also explored by adding electrical noise (white, zero mean) to the simulated radio frequency (RF) signals from which the simulated images were obtained. The resulting signal to noise ratios (SNR) for tissue regions in the B-mode images were approximately 40, 30, 20, 12, 5 and -3 dB, with lower SNR for the vessel regions. In addition, the effect of 2 dB/cm attenuation due to image depth was examined, by comparing results from images with and without attenuation. Images were simulated with different settings (7.0 MHz as centre frequency, 40 MHz sampling frequency, 192 elements, 64 active elements and 64 lines per image, other settings as above), with ellipses using the same eccentricity values as above and φ = {−25◦ , 0◦ , 20◦ }, for a total of 25 images with and 25 without attenuation. For the second step, six different vessels were segmented in three different ultrasound images: one image with the jugular vein and the common carotid artery (560×472 pixels), and two images each with one saphenous vein and arteries (720×480 pixels). The images of the saphenous veins and arteries were acquired from instructional videos [123] of conventional compression ultrasound examinations (used with permission, Davies Publishing). Three experts (one radiologist (Savvas Nicolaou) and two sonographers (Vicki Lessoway and Maureen Kennedy)) segmented all images by tracing the contour using an image editing application. Two of the experts (VL and MK) segmented all images twice, resulting in 5 expert tracings per image. The algorithm was initialized in all cases with N = 50 radii, Q = diag ( 2, 2, 0.1 ), R = 20, M = 5, and x ˆ0|0 = [ rmax /2, rmax /2, 0 ] . Values of rmax ranged from 60 to 80 pixels for simulated images and 70 to 85 pixels for patient images, depending on vessel size. Seed points for each image were determined from the expert tracings, from which a central point was calculated as the mean of 50 evenly sampled points. The expert tracings were then defined 87  Figure 4.3: Location of seed points. Example of seed points (.) generated for a simulated image. Expert tracing shown in gray. in polar coordinates originating at this central point. A new contour was defined by dividing the radial component of each polar coordinate by 4. Pixels that were contained within this new contour were sampled by a factor of 4 and the resulting coordinates, all of which are unique and located roughly at the centre of the vessel, were used as the seed points. Figure 4.3 shows an example of seed points generated for a simulated image. For simulated images, the estimated ellipse parameters were compared to the known parameters and errors were determined as a percentage of the known parameter. The area obtained from the estimated ellipse parameters using πˆ aˆb was compared to the ‘true’ area, calculated from the known parameters using πab. The area from estimated ellipse parameters was also compared to the area resulting from the expert tracing for all images. The expert area was determined by counting the pixels within the traced contours. These area errors are also reported as percentages of the ‘true’ expert areas. Errors between the detected contour and the expert contour were determined using the mean and rms radial distance between boundaries [21]. Given a seed point, along each of the N radii, the expert contour is found and the distance to the detected contour point for the corresponding radii is found. This measurement is used to determine mean (MEr) and rms (RMSEr) errors between true and detected contours. Errors are presented in pixels because scaling factors were known for simulated images (0.0625 to 0.0893 mm/pix), but not for the ultrasound images of patients. However all vessel sizes were similar when measured in pixels, and ranged from ∼100 pixels along their longest axis for simulated images, and ∼80 to ∼150 pixels for patient images. Because the intended application allows for a range of scaling factors, we believe results presented in pixels more accurately represent the performance of our segmentation method. However, image scaling factors should be taken into account when interpreting results for specific cases. The computation time was on the order of 10 to 100 ms for each estimation when tested using Matlab c on a modern PC and unoptimized code, and are adequate for real-time implementation. Typical execution times of the implemented system are presented in Section 4.3.2. 88  Table 4.2: Overall parameter estimation results - simulated images. a b φ FArea EArea MEr RMSEr ◦ (%) (%) ( ) (%) (%) (pix) (pix) √ Thresh. = 1.5 · rmax , (n = 999) (n = 4995) µ -13.68 -2.20 0.80 -15.72 -8.14 -1.06 6.04 σ 8.52 19.89 19.14 16.67 18.03 2.10 2.16 Threshold = 6, (n = 799) (n = 3995) µ -12.44 -2.05 3.82 -14.46 -7.32 -1.00 5.94 σ 7.98 21.38 19.72 17.10 18.85 2.16 2.28 Mean (µ) and std. dev. (σ) of errors for ellipse parameters a, b, and area compared to πab (FArea) and expert tracings (EArea). Values of ellipse parameter φ, and mean MEr and rms RMSEr errors between contour and expert tracing. Results - Simulated Data In the first validation step a total 999 seed points were generated from the expert tracings of the 18 simulated ultrasound images and used as seed points for the algorithm. The number of seed points per image ranged from 23 to 132, and were at most ±13 pixels from the seed in the x direction and at most ±13 pixels from the seed in the y direction. Each detected contour was compared to each of the five expert tracings and estimated ellipse parameters were compared to the known values used to simulate the images. Results are divided into two categories, based on the error threshold defined by (3.1) in Section 3.1.1, used to determine invalid data. The first category uses the value as defined in (3.1), while the second category uses a fixed value of 6 pixels. A valid result was obtained in 100% of evaluations of the algorithm when using (3.1) as the error threshold, and only in 80% of evaluations when using max erRM S = 6 pixels. Overall results are presented in Table 4.2, with mean (µ) errors and standard deviations (σ) for ellipse parameters a, b and the area calculated from πˆ aˆb (FArea), expressed as a percentage of the true known value. The mean and standard deviation of φ is also included. Also shown are µ and σ of the area error when compared to the expert tracings (EArea), as a percentage of the area of the expert tracing. Finally, the mean error (MEr) and the rms error (RMSEr) between the detected contour and expert tracing are presented. The sign indicates over– (+) or underestimation (-). It should be noted that the sample number is 5 times the number of generated seed points for errors obtained when comparing the algorithm results to expert tracings (EArea, MEr, RMSEr), since each detected contour was compared to each of the 5 expert tracings. Table 4.3 presents the mean (µ) and standard deviation (σ) of ellipse parameters and the resulting ellipse area obtained by our algorithm, compared to the true known values. The EArea, MEr and RMSEr errors for individual experts from the simulated images are presented in Tables 4.4 and 4.5. Data has been subdivided as described above, with data obtained when the validation threshold is as in (3.1) in Table 4.4, and when the validation threshold = 6 pixels in Table 4.5. 89  a (pix) True µ  σ  Table 4.3: Detected ellipse parameters. b (pix) φ (◦) FArea (pix2 ) True µ σ True µ σ True µ σ  n  55  51.0  3.98  55  52.44  3.64  0  4.81  31.4  9503  8404  765  256  55  49.79  2.77  36.67  35.63  2.24  0  5.78  17.71  6336  5575  480  172  55  49.53  2.81  27.5  26.85  1.82  0  -1.70  11.34  4752  4189  476  125  55  46.11  2.37  22  20.67  1.17  0  -1.05  8.70  3801  2999  279  100  55  45.71  2.16  18.33  17.63  1.20  0  -2.71  7.44  3167  2537  270  93  55  43.28  1.97  15.71  15.84  4.88  0  -2.93  8.88  2715  2144  597  73  55  44.77  2.82  13.75  13.81  5.15  0  -2.30  8.80  2376  1948  727  69  55  41.77  2.92  12.22  14.64  5.72  0  -4.90  8.44  2111  1919  749  59  55  38.90  3.16  11  9.87  3.00  0  -3.86  5.37  1900  1198  313  52  Mean (µ) and std. dev. (σ) of estimated parameters a, b and φ, and area from πˆ aˆ b (FArea) compared to true values, using threshold = √ 1.5 · rmax .  90  Table 4.4: √ Contour detection results - simulated images, errors per expert, using validation threshold = 1.5 · rmax . SN (n = 999) VL (n = 1998) MK (n = 1998) µ σ µ σ µ σ EArea (%) MEr (pix) RMSEr (pix)  -10.12 -1.28 6.28  17.54 2.14 2.03  -9.19 -1.35 5.91  17.93 2.09 2.14  -6.11 -0.66 6.04  18.19 2.02 2.24  Mean (µ) and std. dev. (σ) of error from area compared to expert tracing (EArea), and mean MEr and rms RMSEr errors between contour and expert tracing, for each expert (SN = Savvas Nicolaou, VL = Vicki Lessoway, MK = Maureen Kennedy).  Table 4.5: Contour detection results - simulated images, errors per expert, using validation threshold = 6. SN (n = 799) VL (n =1598) MK (n = 1598) µ σ µ σ µ σ EArea (%) MEr (pix) RMSEr (pix)  -8.33 -1.06 6.14  17.77 2.16 2.14  -8.78 -1.37 5.84  18.89 2.17 2.26  -5.35 -0.60 5.93  19.17 2.09 2.38  Mean (µ) and std. dev. (σ) of error from area compared to expert tracing (EArea), and mean MEr and rms RMSEr errors between contour and expert tracing, for each expert (SN = Savvas Nicolaou, VL = Vicki Lessoway, MK = Maureen Kennedy).  Table 4.6: Parameter estimation and contour detection results for images with attenuation. Att. a (%) b (%) φ (◦ ) FArea (%) (dB/cm) µ σ µ σ µ σ µ σ – 2  -22.2 -21.7  12.3 11.3  -2.4 -6.7  26.7 18.5  -1.1 0.8  13.0 12.9  -25.1 -27.2  16.6 12.6  Table 4.6 shows algorithm results from 1375 seed points on images with and without attenuation. The error distributions of the data obtained from attenuated images were compared to those from images without attenuation using a two-sample Kolmogorov-Smirnov goodness-of-fit hypothesis test with a significance level of α = 0.05. Parameter estimation errors were found to be significantly different (p-values < 0.0182) from results without attenuation, but with error differences of less than 4%. The results of contour detection and parameter estimation on noisy images are shown in Table 4.7. The same 999 seed points as described above for the noiseless images were used. As with the data from attenuated images, the error distributions from both sets were compared to using the two-sample Kolmogorov-Smirnov goodness-of-fit hypothesis test with a significance level of α = 0.05. While a statistically significant difference was found for the mean contour error, MEr (p-value = 0.0388), and the estimated ellipse angle, φ (p-value = 0.0409), at SNRs of 30 and 12 dB, results remain very similar to those obtained from the noiseless data until the SNR approaches 0 dB. P-values for all measures are less than 0.05 when SNR is -3 dB. 91  Table 4.7: Parameter estimation and contour detection results for images with varying signal-to-noise ratios (SNRs). SNR a b φ FArea EArea MEr RMSEr ◦ (dB) (%) (%) ( ) (%) (%) (pix) (pix) µ σ µ σ µ σ µ σ µ σ µ σ µ σ – 40 30 20 10 5 -3  -13.68 -13.95 -14.06 -13.98 -14.01 -14.73 -18.25  8.52 8.74 8.58 8.58 8.76 9.54 14.31  -2.20 -2.74 -2.53 -2.61 -2.92 -0.54 -7.88  19.89 17.76 17.49 16.68 12.79 27.45 25.29  0.80 -0.77 0.01 -1.39 -0.77 0.91 -1.35  19.14 19.43 19.58 19.47 19.43 19.13 20.42  -15.72 -16.38 -16.37 -16.28 -16.46 -15.64 -23.46  16.67 15.93 14.65 14.99 13.52 20.39 24.16  -8.14 -8.88 -8.83 -8.84 -9.06 -8.15 -17.09  18.03 17.07 16.20 15.01 13.28 21.59 24.93  -1.06 -1.17 -1.16 -1.14 -1.14 -1.13 -2.56  2.10 1.88 1.82 1.90 1.68 2.53 4.61  6.04 6.01 6.03 6.03 5.98 6.28 7.45  2.16 2.00 2.00 2.03 1.90 2.60 4.59  92  Table 4.8: Overall results - patient images. EArea MEr RMSEr (%) (pix) (pix) √ Thresh. = 1.5 · rmax (n = 2820) µ 11.66 1.06 10.47 σ 28.19 5.83 4.53 Threshold = 6, (n = 1835) µ 13.24 1.70 10.09 σ 23.98 5.17 4.30 Mean (µ) and std. dev. (σ) of error from area compared to expert tracing (EArea), and mean MEr and rms RMSEr errors between contour and expert tracing, from patient images ranging from 560×472 to 720×480 pixels. Results - Patient Data The objective of the second validation step was to compare the results of the feature detection algorithm to the expert segmentation in the patient images. Expert tracings were again used to generate a total of 564 seed points. The number of seed points per image ranged from 37 to 144, and were at most ±21 pixels from the seed in the x direction and at most ±12 pixels from the seed in the y direction. Again, each contour was compared to five expert tracings, generating 5n = 2820 total measurements. The mean (MEr) and rms (RMSEr) radial distance between the detected boundaries and the expert boundaries was determined in pixels, and the differences between expert area and the area from the estimated ellipse parameters (EArea) was determined as a percentage of the expert area. These results are presented in Table 4.8. As in the first step, the same two validation thresholds were used for the patient images dataset. Table 4.10 shows errors for individual experts when the validation threshold is as in (3.1) with a 100% validation rate, and Table 4.11 shows errors for individual experts when the validation threshold = 6 pixels with a 65% validation rate. Examples of the results of the feature detection algorithm and parameter estimation on human data are presented in Figure 4.4. The mean (µ) area values for the patient images are also presented in Table 4.9, alongside the standard deviation (σ), the coefficient of variation V (ratio of standard deviation to mean), and the ‘True’ expert values, obtained when using validation threshold = 6 pixels. The patient images are shown in Figure 4.4. It is observed that the smallest V was obtained from a more elliptically shaped vessel (Figure 4.4(c)). The largest V was obtained from the image with the smallest feature (Figure 4.4(f)), which may explain the large variation. While µ is close to the true value, σ for this image was the largest of all.  93  (a)  (b)  (c)  (d)  (e)  (f)  Figure 4.4: Patient data (images cropped to 200 × 200 pixels) with examples of detected contours (white circles) and estimated ellipses. The estimated semi-major axis a and semi-minor axis b are presented as a solid and dashed line, respectively, oriented according to the estimated φ. Expert tracings presented in green. (a), (b), (d) and (f) reprinted with permission [123], (c) and (e) obtained from laboratory tests. Even though expert tracings are not elliptical, the detected contours correspond to the vessel shapes illustrating that the contour does not have to be elliptical even though an ellipse model is used.  Table 4.9: Overall results, in pixels - patient images. EArea (pix2 ) Image n True µ σ V (Figure 4.4) 4862 7113 7172 6546 6240 2514  6488 8532 6782 8528 6731 2113  1572 435 272 810 1020 1697  0.24 0.05 0.04 0.09 0.15 0.80  135 420 495 360 380 45  (a) (b) (c) (d) (e) (f)  Mean (µ), std. dev. (σ) and coefficient of variation (V ) of estimated area (EArea) compared to true value from expert tracing, using validation threshold = 6.  94  Table √ 4.10: Contour detection results - patient images, errors per expert, using validation threshold = 1.5 · rmax . SN (n = 564) VL (n =1128) MK (n = 1128) µ σ µ σ µ σ EArea (%) 5.95 27.26 7.62 25.72 18.55 29.58 MEr (pix) 0.64 6.13 0.31 5.74 2.03 5.63 RMSEr (pix) 11.03 4.82 10.52 4.58 10.13 4.29 Mean (µ) and std. dev. (σ) of error from area compared to expert tracing (EArea), and mean MEr and rms RMSEr errors between contour and expert tracing, for each expert (SN = Savvas Nicolaou, VL = Vicki Lessoway, MK = Maureen Kennedy).  Table 4.11: Contour detection results - Patient images, errors per expert, using validation threshold = 6. SN (n = 367) VL (n =734) MK (n = 734) µ σ µ σ µ σ EArea (%) 9.14 24.20 9.45 21.07 19.08 25.40 MEr (pix) 1.47 5.55 0.97 4.92 2.56 5.10 RMSEr (pix) 10.69 4.59 10.16 4.28 9.72 4.14 Mean (µ) and std. dev. (σ) of error from area compared to expert tracing (EArea), and mean MEr and rms RMSEr errors between contour and expert tracing, for each expert (SN = Savvas Nicolaou, VL = Vicki Lessoway, MK = Maureen Kennedy).  4.1.2  5–Parameter Ellipse Model  Convergence Tests The contour detection algorithms using a 5–parameter ellipse model were evaluated to determine how well parameters were estimated, similar to what was done for the 3–parameter ellipse model algorithms in Section 4.1.1. Random ellipse parameter values within a predetermined range were generated and substituted into the 5–parameter ellipse output equation (3.16) with the known θk value in order to generate a measurement v+ k =  xe P v ±  (xe P v)2 − (v P v) xe P xe − 1 v P v  (4.3)  where v + k refers to the real, positive solution as described in Section 3.1.3. A measurement residual is then calculated z k+1 = v + xk+1|k ) k+1 − Ck (ˆ  (4.4)  using this output equation evaluated with the predicted state x ˆk+1|k . As before the semi–axes values, a and b, where chosen in the range [10, 100]15 with a uniform distribution and the angle φ was chosen in the range [−π/4, π/4] ([−45◦ , 45◦ ]), also with a uniform distribution. Finally, the minimum value of mina,b = [a, b] was calculated to generate the ellipse 15  Arbitrary units were used, however the range in values is similar to what is encountered in a clinical setting.  95  Table 4.12: Convergence benchmarks for 5–parameter ellipse models. Shows the average number of iterations required to reach each benchmark. Output n Benchmark 1 Benchmark 2 Benchmark 3 Algorithm (no.of iterations) (no.of iterations) (no.of iterations) Small step  1758  2.98  3.97  5.52  Large step  1838  1.54  1.77  2.09  centre offset, which was in the range [−mina,b /2, mina,b /2] with uniform distribution. This ellipse centre offset was calculated in this manner so that the origin would always lie inside the ellipse. Otherwise, the solution to (3.16) would yield imaginary results. The initial conditions for the algorithms were [50 50 0 0 0] . The same three benchmarks used in Section 4.1.1 were set, and criteria was added for the ellipse centre estimation. The benchmarks were 1. When the error for the ellipse semi–axes was less than 10% of the true parameter value, the angle φ error was less than 5◦ and the error for the ellipse centre coordinates was less than 3 ‘pixels’. 2. When the error for the ellipse semi–axes was less than 5% of the true parameter value, the angle φ error was less than 2◦ and the error for the ellipse centre coordinates was less than 1 ‘pixel’. 3. When the error for the ellipse semi–axes was less than 1% of the true parameter value, the angle φ error was less than 1◦ and the error for the ellipse centre coordinates was less than 0.5 ‘pixels’. The number of iterations to reach each benchmark was recorded. The error between the estimated and true parameters were calculated at the end of each iteration and the errors for the ellipse semi–axes were also normalized by the true parameter values. In addition the mean squared error (MSE), as well as the root-mean squared error (RMS), between the detected ‘contour’ and the true contour were calculated. The algorithms were executed 2000 times with a maximum of 20 iterations each. If the three benchmarks were not reached within the maximum number of iterations, then it was considered that the algorithm did not converge. The average number of iterations for each of the benchmarks for two algorithms are reported in Table 4.12. These algorithms are a 5–parameter ellipse model using the output approach, and the same 5–parameter ellipse model using the staggered scanning or large-step approach described in Section 3.1.4. The number of times the algorithms converged is also shown (n) in the Table.  Table 4.13: Standard deviation of errors for 5–parameter ellipse models. Output a b φ xe ye Algorithm Small step  0.420  0.411  0.227  0.141  0.139  Large step  0.224  0.223  0.135  0.097  0.095 96  (a) Ellipse Parameter Errors for a and b  (b) Ellipse Parameter Errors for φ  (c) Ellipse Parameter Errors for xe and ye  Figure 4.5: Final ellipse parameter errors obtained during convergence tests for the 5–parameter ellipse model using the output approach.  (a) Mean Squared Error  (b) Root Mean Squared Error  Figure 4.6: Final contour errors obtained during convergence tests for the 5–parameter ellipse model using the output approach. The results for both models are good. When using the small step output model, convergence was obtained in 1758 of 2000 cases (87.9%) and convergence was obtained in 1838 of 2000 cases (91.9%) when using the large-step or staggered scanning approach. All ellipse parameter values are within the expected ranges, as can be seen in Figures 4.5 and 4.7. It should be noted that the number of iterations needed to reach each benchmark using a large step is quite smaller than if using the small angular step. The main reason for the large step approach is to speed up parameter convergence, and this goal has been achieved. An added benefit of this is that spread of the errors of the estimated parameters is also reduced when using a large angular step. This reduced deviation can be seen when comparing Figures 4.5 and 4.7. The standard deviation of the ellipse parameter errors for the two methods is also presented in Table 4.13.  97  (a) Ellipse Parameter Errors for a and b  (b) Ellipse Parameter Errors for φ  (c) Ellipse Parameter Errors for xe and ye  Figure 4.7: Final ellipse parameter errors obtained during convergence tests for the 5–parameter ellipse model using the output approach and large step size.  (a) Mean Squared Error  (b) Root Mean Squared Error  Figure 4.8: Final contour errors obtained during convergence tests for the 5–parameter ellipse model using the output approach and large step size. Full error range not shown. However, the use of a large step significantly increases the error in the detected contours as illustrated in Figure 4.8. The errors obtained when using a small step are presented in Figure 4.6. The main reason for this drastic increase is since parameters converge very quickly, often in just one iteration, any contour errors at the beginning of the estimation (when the parameter values have yet to be determined) will contribute to the final errors. This is illustrated in the example in Figure 4.9. Notice how there are large errors between the estimated contour and the true location. These large errors correspond to the radii that were first estimated (the radii for the contour were re-ordered in the graph).  98  Figure 4.9: Example of parameter convergence using a large angular step and 64 radii. Values of estimated contour and parameters (blue) are shown along with true values (black). Notice the large contour errors at several locations, which correspond to the radii that were first estimated (contour radii have been re-ordered). One solution to reduce the contour error is to estimate ellipse parameters using a large angular step in one iteration, and then refine and smooth out the contour using a small angular step, with the first results as initial conditions. Similar results as those just presented are expected if the large step approach were applied to the 3–parameter ellipse models. Simulated and Patient Image Data - Procedure The performance of the contour detection algorithms using a 5–parameter ellipse model were also evaluated on simulated and patient images, as was done for the 3–parameter ellipse based methods in Section 4.1.1. As before, the simulated images were used to determine the accuracy of the 99  Table 4.14: Errors between estimated ellipse parameters and true values.  3PAR 5PAR 5PAR-s 5PAR-l  a ˆ (%) µ σ -10.7 16.3 -6.6 7.9 -16.6 14.2 -4.9 7.4  ˆb (%) µ σ 1.1 16.4 2.2 3.3 4.3 19.9 3.5 5.1  φˆ (◦ ) µ 4.9 -0.1 0.9 -2.4  σ 25.3 3.5 22.3 8.2  x ˆe (pix) µ σ – – -0.69 3.27 0.99 9.73 -1.19 2.88  yˆe (pix) µ σ – – -0.40 0.65 0.48 1.73 0.46 0.87  algorithm for detecting ellipse parameters and contours from elliptical features with known characteristics, and the patient images were used to determine the accuracy with which the algorithm detected the actual feature contour determined by an expert tracing. The same noise-less simulated images generated using Field II c [80] with elliptical features with known parameters from Section 4.1.1 were used in the first step, and the same six vessels in three different ultrasound images were used for the second step. The effect of image noise on the accuracy of the algorithm was not evaluated for the 5–parameter ellipse algorithms. In addition, the same 5 expert tracings were used. Initial conditions for the algorithms were the same for all images, except for the radial search area rmax (depending on vessel size) and the initial seed points. Seed points were determined from expert tracings as before, except that in these experiments the seeds were located roughly in a ring around the vessel centre, and only used once per image. No seeds points at the centre of the vessel were used. In all cases, N = 64 radii (M = 8 for the staggered scanning approach), R = 10, , the state covariance P0 was initialized to the identity matrix and Q = diag ( 2, 2, 0.1, 2, 2 ) x ˆ0|0 = [ rmax /2, rmax /2, 0, 0, 0 ]  .  (4.5)  Values of rmax ranged from 60 to 80 pixels for simulated images and 70 to 85 pixels for patient images, depending on vessel size. Estimated parameters were compared to known parameters in simulated images, and expressed as a percentage of the known parameter. Error distributions were approximated using a Gaussian with mean µ and standard deviation σ. For all images, the detected contour was compared to expert tracings, and the mean (MEr) and rms (RMSEr) radial distance between boundaries was calculated [21]. The area error (AE) between the ellipse from πˆ aˆb and ‘true’ area enclosed by expert segmentation was found, and the overlapping areas from both detected contour (AOC) and estimated ellipse (AOE) when compared to expert tracings were calculated. Execution time was also measured. As a comparison, these experiments were also performed on the 3–parameter ellipse model using the same initial settings just described.  100  (a) Simulated Image  (b) Patient Image  Figure 4.10: Detection results on simulated (a) and patient (b) data with contour ‘+’ and ellipse ’– –’ using a 3–parameter ellipse algorithm, and contour ‘o’ and ellipse ’–’ using a 5–parameter ellipse algorithm, both with the same initial seed point. The results of the 3–parameter algorithm are clearly not adequate, while the results from the 5–parameter version are quite good. Simulated and Patient Image Data - Results The 3–parameter ellipse model from Section 3.1.2 using the output approach was run using N = 64 and 3 radial traversals (3PAR). The 5–parameter ellipse model (5PAR) was also run with 3 radial traversals. In addition, the 5–parameter model was also evaluated using only 1 traversal, once using a small step size (5PAR-s) and once using a large step size (5PAR-l), as described in Section 3.1.4. All algorithms were executed 855 times on the simulated images and 485 times on the patient images. Since there are no constraints on the state, it is possible that the estimates in the 5– parameter model yield a result with xc outside the ellipse, in which case imaginary values for ρ1,2 are obtained and the contour detection is aborted. For each algorithm (5PAR, 5PAR-s, 5PAR-l) this occurred respectively 18, 8 and 6 times for simulated images and 5, 3 and 5 for patient images - these results were discarded. Parameter estimation results are presented in Table 4.14, while the mean (MEr) and rms (RMSEr) contour errors, as well as the ellipse area error (AE) and area overlap, with mean (µ) and false positive (FP) and negative (FN) values, for ellipse (AOE) and contour area (AOC) are presented in Tables 4.16 and 4.17, for simulated and patient images, respectively. Figure 4.11 shows the error distributions for MEr and AE, for both simulated and patient images. Mean execution times with unoptimized code are shown in Table 4.15. Finally, examples of contour detection are presented in Figures 4.10(a) and 4.10(b), and the advantages of the 5 parameter ellipse model (5PAR) can be observed, as there are large errors for the 3 parameter model. Algorithms were initialized using the same seed (‘*’). 101  (a) Mean Contour Errors  (b) Mean Contour Errors  (c) Ellipse Area Error  (d) Ellipse Area Error  Figure 4.11: Mean contour errors (pix) in simulated (a) and patient (b) images, and ellipse area errors (%) in simulated (c) and patient (d) images, for 3 (3PAR) and 5 (5PAR) parameter ellipse models, and a 5 parameter model with large step size (5PAR-l).  Table 4.15: Algorithm execution times (ms). 3PAR 5PAR 5PAR-s Simulated Images 157.13 280.07 114.41 Patient Images 102.49 342.53 113.10  5PAR-l 103.98 126.42  102  Table 4.16: Errors between detected contour and MEr (pix) RMSEr (pix) AE (%) µ µ µ σ σ σ 3PAR 2.31 3.29 11.82 3.86 3.1 23.7 5PAR 1.47 0.95 3.69 1.02 4.9 4.6 5PAR-s 1.00 4.78 9.99 2.08 -5.6 7.4 5PAR-l 1.82 1.25 4.98 1.78 8.1 6.8  expert tracings for simulated images. AOE (%) AOC (%) µ µ FN FP FN FP 74.6 25.3 28.5 81.3 18.6 23.3 99.8 0.1 7.8 95.5 4.5 9.3 85.2 14.8 24.4 77.8 22.1 31.6 94.0 5.9 10.7 79.1 20.8 25.4  Table 4.17: Errors between detected contour and MEr (pix) RMSEr (pix) AE (%) µ µ µ σ σ σ 3PAR 4.38 4.81 15.84 4.44 11.7 24.4 5PAR 3.64 3.64 9.67 6.38 16.0 17.7 5PAR-s 0.72 2.72 11.82 3.89 1.3 12.8 5PAR-l 3.59 3.38 11.60 5.20 19.7 21.1  expert tracings for patient images. AOE (%) AOC (%) µ µ FN FP FN FP 78.5 21.4 32.9 81.8 18.2 29.6 90.5 9.5 15.2 95.9 4.0 10.4 84.7 15.3 15.5 78.2 21.7 24.0 89.8 10.2 25.4 82.6 17.4 32.3  103  (a) Ellipsoid Parameter Errors for (b) Ellipsoid Parameter Errors for rz (c) Ellipsoid Parameter Errors for a, b and c xe , ye and ze  Figure 4.12: Final ellipsoid parameter errors obtained during convergence tests for the ellipsoid model using the output approach. In these experiments, the seed points were chosen to accentuate the strengths of the 5– parameter models. When all seeds are at the centre of the vessel, the results from the 3– and 5– parameter models are very similar. Indeed, when xe = xc the 5 parameter model is reduced to the 3 parameter model with a single ρ. The data clearly show that using 5PAR results in good parameter estimation and smaller error variations when compared to the 3PAR algorithm, as seen by the generally smaller σ in the Tables. While the ellipse area (AE) from 5PAR overestimates the expert tracings in Tables 4.16 and 4.17, correspondence is good as seen by AOC and AOE. In addition, an estimate of the ellipse centre xe is only obtained when using a 5 parameter model. When the number of traversals is reduced (5PAR-s), the parameter errors have larger variations, and area correspondence decreases as seen from AOE and AOC values. This is due to the slow and incomplete parameter convergence explored above. The advantage of using a larger step size (5PAR-l) can be seen from the decreased variation in the parameter errors and mean errors similar to using more iterations (5PAR). Also, the ellipse area (AE) and ellipse correspondence with expert tracing (AOE) with this method is comparable to that obtained from 3 traversals. However, the detected contour is not as accurate as seen from larger RMSEr and lower AOC values. For applications where only an estimate of the transverse area is necessary and not an exact contour, using the larger step size and fewer iterations provides an adequate solution. While the reported execution time indicates the 5–parameter algorithms take longer than the 3–parameter model, real-time implementation (20 Hz) should not pose significant problems. The optimized version of the 3–parameter ellipse model algorithm takes about 20 ms per image [66], and the 5–parameter model has a similar potential.  4.2  Vessel Tracking  The vessel tracking with delayed measurements presented in Section 3.2.1 on page 46 was implemented, including seed correction using the SSA approach from Section 3.2.6 on page 55, and tested as follows. Three representative series of ultrasound images of 100 images each, where a 104  Table 4.18: Errors and execution time for different SSA sizes. Err. S1 (pix) Err. S2 (pix) Exec. Time (s) µ σ µ σ µ σ Full mask 4.539 3.423 3.082 1.574 0.388 0.049 Half mask 4.595 3.383 3.059 1.599 0.224 0.013 Fifth mask 4.625 3.435 3.010 1.705 0.082 0.003 Mean (µ) and std. dev. (σ) of errors from tracking a vessel centre in a slow (S1) and a medium (S2) step images series when using all, half, and one-fifth the pixels in a SSA mask. Execution time is also shown. feature from an ultrasound phantom was seen to move from left to right (a slow (S1), medium (S2) and fast (S3) step) were acquired and saved for offline evaluation, along with corresponding location data. The peak speeds were 1.6, 6.9 and 12.6 pixel/frame (1.4, 6.2, 11.2 mm/sec @ 10 Hz, respectively) along the x axis for the slow, medium and fast steps, respectively. Movement along the y axis was considered negligible. The centre of the features was determined manually, and the tracking algorithm was used to estimate the vessel centres using image and sensor data. An error measure was determined for each image by calculating the distance from the tracked position to the true segmented position. The tracking method presented in eqs. (1) and (2) in [3] was also implemented and used as a baseline. This system also uses a constant velocity model, but the measurement consists of the current seed location obtained by projecting the previous centroid onto the current image frame. The state estimate is calculated in a single global reference frame. The estimated seed point was also corrected using seed correction using SSA from Section 3.2.6. It was found that this system does not work sufficiently well, as the seed location was repeatedly lost especially for faster movements, through the implementation of this model on a prototype of the DVT system from [66], and motivated the current tracking model. The use of the SSA approach was also investigated by generating estimates of the vessel locations using the tracking algorithm with delayed measurements from Section 3.2.1 on page 46 and three different settings for the SSA as described on page 55 in Section 3.2.6. Errors between the true positions and tracked positions were calculated as above for each of the SSA settings. A two-tailed t-test was used to determine whether there was any significant difference between the mean errors obtained from the tracked positions when using the different SSA settings. The tracking was implemented using an 80 pixel SSA, with 1/5 pixels within the mask (see Figure 3.11 on page 55). The method from [3] was implemented using an 80 pixel mask. The tracking results with errors are presented in Figures 4.13 and 4.14. For slower movements, both implemented tracking methods are seen to work appropriately (Figure 4.13), but for larger movements (Figures 4.14(a) and 4.14(b)), only our method with delayed location measurements can track the seed location, as shown by the data. For the SSA validation, the slow (S1) and medium (S2) step ultrasound image series were used, and the algorithm was implemented using a 80 pixel square mask, using all, 1/2 and 1/5 the pixels within the mask. Typical results are presented in Figure 4.15. The errors between the tracked 105  (a) Slow Step  (b) Error Slow Step  Figure 4.13: Seed tracking results. A seed point was tracked in a slow step (a), and a medium and fast step (Figure 4.14). Manual segmentation is represented by the solid line, the previous tracking method by (+) and the method presented herein by (.), and the x (above) and y (below) coordinates are presented. Errors (b) are also shown. points and the true location are presented in Table 4.18. No statistically significant difference between the SSA methods was found, while a reduction in execution time proportional to the number of pixels used can be clearly observed. The main drawback to this tracking implementation is the possibility of introducing more noise into the system. In addition to any noise introduced because of the electromagnetic sensor used for location measurements, the sensor location on the ultrasound probe is far from the actual image plane location. Therefore, sensor errors may be compounded leading to inaccurate tracking. The delayed measurement models from Section 3.2.1 and the velocity models from Section B.2.1 on page 171 were qualitatively evaluated and it was found that the model with the velocity in the state was susceptible to small variations in sensor readings. On the other hand, the constant velocity model with delayed measurements which includes location information works very well for small and large seed point movements. Any deterioration because of location measurement noise is outweighed by the possibility of tracking larger movements.  4.2.1  Tracking Vessel Bifurcations  The method of thresholding and labeling the flow data for identifying vessel bifurcations was evaluated. It is determined whether centroids of the flow regions found from thresholded flow images are contained within a detected contour. If flow is found outside this contour, then a vessel bifurcation has been found. This can be considered a single image approach, since it can be determined whether there are one or more vessels in the image area using one frame (using both 106  (a) Medium Step  (b) Fast Step  (c) Error Medium Step  (d) Error Fast Step  Figure 4.14: Seed tracking results. A seed point was tracked in a slow (Figure 4.13), medium (a) and fast step (b). Manual segmentation is represented by the solid line, the previous tracking method by (+) and the method presented herein by (.), and the x (above) and y (below) coordinates are presented. Errors for the medium (c) and fast (d) steps are also shown. colour flow and B-mode data). Five image series, ranging from 90 to 227 images each, and containing scans of the carotid artery with bifurcations (5 image series with 193, 200, 150, 107, 100 images, respectively) were used in the evaluation. These were scans of vessels viewed in transverse axis obtained while moving the probe along the long axis of the vessel. Flow information was available for all images. Flow data was thresholded using values greater than 130 and less than 124 (on a scale of 0 to 255), and labeled according to the flow direction. A second pass using thresholds of values greater than 132 and less than 122 was also done on all images. Only regions greater than 200 pixels were used. An initial seed point was selected for the first image in each series and tracked in subsequent images, while calculating a contour for each image. It was determined whether thresholded regions were inside or outside the detected contour, and the direction of flow inside the contour was determined. If a flow region had the same direction and was found outside the detected contour, a new contour was generated at that location. The first image in each series was not used in the evaluation. All newly generated contours were classified as correct or incorrect, based on whether the tracked vessel had split or not. The vessel bifurcation was defined as the first frame in each image series where flow could be clearly seen throughout the vessel lumen, and the vessel cross–section was clearly increased, but not yet split into two vessels. 107  (a) Tracking Slow Step  (b) Execution Time Slow Step  (d) Tracking Medium Step  (e) Execution Time Medium Step  (c) Error Slow Step  (f) Error Medium Step  Figure 4.15: SSA validation. A seed point was tracked in a slow (a) and medium step (d), using all (+), half (.), and one fifth the pixels (x). Execution times per image for the slow (b) and medium step (d) are presented, as well as respective errors (c), (f). For each frame, the author determined whether any additional contours generated using this method were correctly identified, by recording the number of cases where no new contours were correctly added (true negatives, TN), where new contours were correctly added (true positives, TP), where new contours were incorrectly added (false positives, FP), and where new contours were not added but should have been (false negatives, FN). An additional evaluation was done taking into account the consecutive nature of the images in the series. It was assumed that once a new contour is generated for a vessel bifurcation, the contour detection and seed tracking take over. Therefore, any false negatives encountered in the initial evaluation after the first correctly identified contour were not counted and assumed to be identified correctly. The number of incorrect contours (false positives) remains the same. A total of 745 images were evaluated, of which 337 showed a bifurcating vessel. The results are summarized in Table 4.19. Results when only using a single image to determine if a vessel has split are presented, as well as those that take into account the image series, in effect converting this approach into a method that uses multiple images to determine a vessel bifurcation. These results are closer to what would be expected in a real-time implementation. When using a low threshold, the main reasons for missed vessel bifurcation detection include no flow in the bifurcation (188, 82.5%), color flow aliasing (17, 7.5%) and both vessels consisting of a single thresholded region (15, 6.6%). Assuming that contour detection and seed tracking, and therefore using multiple images, is used for detecting bifurcations, having no flow in the bifurcation in some images can be directly addressed. This is also reflected in the data in Table 4.19. 108  Table 4.19: Detecting vessel bifurcations using flow data, first data set Correct, TP Missed, FN Incorrect, FP cnt.  %  cnt.  %  cnt.  %  Low threshold (130 & 124) – single image  109  32.3%  228  67.7%  66  16.2%  Low threshold (130 & 124) – multiple image  309  91.7%  28  8.3%  66  16.2%  High threshold (132 & 122) – single image  83  24.6%  254  75.4%  43  10.5%  High threshold (132 & 122) – multiple image  252  74.8%  85  25.2%  43  10.5%  Table 4.20: Detecting vessel bifurcations using flow data, second data set Correct, TP Missed, FN Incorrect, FP cnt.  %  cnt.  %  cnt.  %  Low threshold (130 & 124) – single image  236  60.0%  157  39.9%  0  0.0%  Low threshold (130 & 124) – multiple image  385  97.9%  8  2.0%  0  0.0%  The reasons for an incorrect contour (false positives) were color flow aliasing (24, 36.4%), other vessels in the region of interest (22, 33.3%), and the flow region extending outside of the detected contour or a single flow signal split into multiple regions (16, 24.2%). Results where similar when using the higher threshold. This method was evaluated on three additional images series of a carotid artery, consisting of 170, 250 and 180 images respectively. Unlike the previous image series, the ultrasound flow settings were selected to best suit the vessel bifurcation detection method. This consisted mainly of setting a high persistence for the flow data, and carefully selecting flow velocity to avoid color aliasing artifacts. All other settings were as described above, using a low threshold. A total of 597 images were evaluated, of which 393 showed a bifurcating vessel. The results are summarized in Table 4.20. The biggest difference is the elimination of incorrect contours or false positives. Also, there is a significant improvement of correct bifurcation detection, even when only using a single image. The main reasons for missing a vessel bifurcation are no flow data in the bifurcation (97, 62%) and both vessels consisting of a single thresholded region (41, 26.1%). As before, when including information from multiple images, this method provides a much improved detection of vessel bifurcations. It is clear from these results that when flow settings are selected appropriately, better results are obtained. Even so satisfactory results can be obtained using typical flow data. This approach has also been implemented in real–time, with satisfactory results. In addition, this method can be used as an automatic seed initialization for both types of vessels (veins and arteries), by using the centroid of the detected areas as seed points. 109  (a) Venous Compression Data  (b) Arterial Compression Data  Figure 4.16: Typical human compression data consisting of normalized transverse area vs. normalized force is presented. Vein (a) transverse area almost disappears as force is applied, while artery (b) transverse area does not change over range of applied force.  4.3 4.3.1  DVT Screening System Receiver Operating Characteristics (ROC) Curves for TAR and Slope Criteria  The system underwent initial laboratory testing on healthy volunteers (n = 3), where several vessel segments belonging to the deep venous system were scanned in each case (total n = 10). Typical results of a compression exam on superficial femoral vein are presented in Figure 4.16(a). The results show a line fit to the data with slope near -1 and a small TAR value indicating the absence of DVT as expected. An arterial segment was also examined, representing an incompressible vessel. The results for one compression-release cycle are presented in Figure 4.16(b), and show a high TAR value and slope value near 0 clearly indicating an incompressible vessel, as expected. These results were initially reported in [60]. Afterward, the TAR and slope data generated by the system from healthy volunteers was used to calculate sensitivity and specificity values for each individual cross-section. Veins were considered ‘healthy’ (n = 41) and arteries were considered ‘diseased’ (n = 15). Receiver operating characteristic (ROC) curves were generated for different cutoff values, and presented in Figure 4.17. Optimal cutoff value for the TAR as determined by the likelihood ratio (LR) was 55% with a sensitivity of 100% and specificity of 97%, while the optimal cutoff value for the slope was -0.2 with a sensitivity of 60% and specificity of 97%. Both of the ROC curves for the TAR and slope criteria indicate that these measures are quite useful, as there is a large area left of the diagonal in Figure 4.17.  110  (a) ROC for TAR  (b) ROC for slope  Figure 4.17: ROC curves for TAR (a) and slope (b) venous compressibility criterion.  4.3.2  System Evaluation  Typical execution times (in milliseconds) for the system used in the clinical pilot study (details in Section 4.3.3) are presented in Table 4.21. Data was obtained from scanning one vessel (n = 2625 frames processed). This implementation of the system operated on average at about 16Hz. Maximum execution time occurs during the model building and compression exam procedure, after all compression data has been acquired for a single location, and the compressibility measures are being calculated and acquired data is written to disk. In this case, this occurs in 3 of 2625 cycles (3 compression exams for this model).  4.3.3  Clinical Pilot Study  A pilot study was developed for an initial evaluation of a prototype DVT screening system and to provide data for system parameters. An additional objective was to obtain information about the patient population to assist in the design of further, more complete, clinical studies. A protocol for the study was presented to and approved by the appropriate ethics review boards (approval forms can be found in Appendix E). The prototype system consisted of an ultrasound machine (RP500, Ultrasonix Medical Corp.), the sensorized ultrasound probe, and the image processing algorithms that detect and track the vessel contours that was described in Section 2.2.3.  Table 4.21: Execution time (ms) - DVT screening system. Read  Contour/  Model/  Render/  Total  Sensors  Tracking  Comp.  Display  Time  Mean  1.2  23.4  10.0  10.7  62.1  Max  22.8  96.8  4,281  36.4  4,342  Min  0.2  0.01  0.01  1.2  11.9 111  Figure 4.18: DVT system interface. The virtual vessel model is presented on the left, while ultrasound data with detected contour is displayed on the right. Figure 4.18 shows the DVT system interface. The ultrasound image is displayed on the right with the detected contour, while the 3-D vessel model made from the contours is shown on the left, with compression data mapped to the surface using colour. A virtual plane indicates the location of the current ultrasound image with respect to the model. Study Protocol The DVT screening system was used to scan segments of the superficial femoral vein and assess them for compression. Vessel bifurcations were not scanned. When possible, more than one vessel segment per patient was scanned. A registered ultrasound technician performed all examinations. The scanning protocol consisted of an initial scan to build the 3-D model with the detected contours, followed by a second scan to assess the vessel for compression. Compression data was gathered in compression-release cycles, which consist of having the examiner press down with the ultrasound probe at one location, and then release. Compression data was assigned to the closest model contour or slice. The system indicated which contours of the vessel had compression data, or insufficient or no data at all by changing the colour at the location of that contour. The model in Figure 4.18 has a slice with no compression data, as indicated by the blue section of the vessel model. If the system failed to generate an adequate model, as determined by the examiner, a new model was generated. An examination was considered complete when compression data was acquired for the entire vessel model. An examination could be concluded at the discretion of the examiner because of concerns for the patient or technical difficulties. The patient could request to stop the examination at any time. All individual images, sensor measurements, detected contours and all individual compression assessments used for the 3-D model were saved, as well as the 3-D model itself. The study took place at the University of British Columbia Hospital, part of the Vancouver Hospital & Health Sciences Centre. Patients scheduled for conventional compression ultrasound examinations were asked to participate, and only those who could provide informed consent and 112  agreed to participate were scanned. Typically, patients were screened using the DVT system after the conventional examination. The results from the conventional ultrasound examination were used as the gold standard. It must be noted however that while compression ultrasound itself has undergone rigorous testing throughout the years to establish sensitivity and specificity, errors may still arise. In other words, there is still some degree of separation between the actual truth (presence or absence of thrombus) and what CUS detects. By using the CUS results as the gold standard, we are limited in that we can only know whether our system is as good as CUS. Results Twenty-three patients (9 male, 14 female) scanned between June 20, 2005 to February 7, 2006, with an average age of 59.3 years (min 22, max 90). Patients were classified as having light (7), medium (13), or heavy (3) build. The reasons for undergoing the conventional DVT examination were pain in 12 cases, swelling in 12 cases, prolonged immobilization in 2, or another in 12 cases. Of these, recent surgery and recent flight accounted for 3 cases each, while the remainder listed various causes. Three patients had a known history of DVT. A total of 41 vessels were scanned on 53 attempts using the DVT screening system, of which 15 (8 patients) were from left limbs and the remaining 26 (16 patients) were right limbs. The mean scan time per vessel was 143.44 s (2.39 min) with a standard deviation of 60.35 s (1.01 min), and a minimum of 69.56 s (1.16 min) and maximum of 442.54 s (7.38 min). Results from the conventional CUS examination were used as the gold standard. From these, 2 patients were found positive for deep vein thrombosis, corresponding to 3 vessel scans. A similar prevalence of DVT of 20-30% has been reported elsewhere [53]. Three-dimensional models of the scanned vessels were constructed in all 41 cases. An average of 6.51 contours (min 3, max 9) were used to construct each model. There was a total of 41 models and 266 model slices and only one model averaged two images to create one slice. Average model length (long axis), determined by measuring the distance between the contours or slices on opposite ends, was 4.89 cm with a standard deviation of 1.41 cm, and a maximum and minimum model length of 8.19 cm and 1.59 cm, respectively. An average of 7.9 compression assessments (min 2, max 15) were performed per vessel model or equivalently 1.22 compression assessments (min 0, max 6) per model slice. Each compression assessment consisted of an average of 10.62 images and sensor measurements (min 1, max 28). A total of 14 models contained at least one model slice without any compression data. Of these, 7 had an empty slice at one end, and 6 had empty slices somewhere in the middle, totaling 25 slices (9.4%) without compression data. There were 185 slices (69.8%) with one compression assessment, while the remaining 55 (20.7%) had two or more compression assessments. Tables 4.22 and 4.23 show the minimum and maximum applied forces during the compressionrelease cycles. Data is also separated based on patient build. TAR and slope values were calculated from each compression assessment, generating an individual value for each model slice. From these, overall TAR and slope values were generated in several manners, such as using the maximum TAR value calculated from all the slices in a single model, or the mean slope value obtained from the same model. The data reported here is obtained 113  Table 4.22: Minimum applied force during compression exams. Patient Mean Std. Min Max n Build (N) Dev. (N) (N) (N) Light Medium Heavy  1.51 0.93 1.11  1.46 1.10 0.87  0 0 0  4.19 3.97 2.83  68 137 35  Overall  1.12  1.21  0  4.19  240  Table 4.23: Maximum applied force during compression exams. Patient Mean Std. Min Max n Build (N) Dev. (N) (N) (N) Light Medium Heavy  18.04 16.15 12.29  7.06 6.31 4.15  6.61 5.62 5.09  32.69 32.57 22.12  68 137 35  Overall  16.12  6.5  5.09  32.69  240  by taking the maximum TAR and maximum slope values for each vessel model, from which the statistics presented in the table are calculated. That is, for all the compression assessments done for a vessel segment, the maximum TAR and slope is used. The resulting TAR and slope values for healthy and diseased vessels are presented in Table 4.24. The system sensitivity and specificity using the TAR were 33.3% and 92.1%, respectively, while the sensitivity and specificity when using the slope measure were in the range of 50–89.5% and 66.7–100%, respectively. The range of values is obtained from the different manners in which the compressiblity measures are calculatd for each scanned vessel, as just described above, by using mean or maximum values of each measure. This is discussed in detail in Section 5.3.2. The force data from the compression data shows interesting results. The mean minimum force is about 1 N, as expected for an uncompressed vessel. The average maximum applied force is about 16 N, but varies depending on patient build. It would be expected that heavier patients would require more force applied to compress the vessel, but the data indicates otherwise. Indeed, there is a significant difference in maximum applied force between light and medium patients (ttest, p-value = 0.053219, α = 0.10), between light and heavy patients (t-test, p-value = 0.000023, α = 0.05) and between medium and heavy patients (t-test, p-value = 0.000756, α = 0.05). It is believed that this difference arises from the different physical properties of fat and muscle tissue.  4.4  Usability Evaluation of the DVT Screening System  There are two main goals for a DVT screening system as described in this dissertation. The first concerns the functionality of the system, or whether the system can in fact detect DVT with the required clinical accuracy. The second concerns the need for the system to be used in a fast manner, by a range of potential users. This section deals with the latter. 114  Table 4.24: Compression assessment results. Mean Std. Dev. Min Max TAR Healthy TAR Diseased Slope Healthy Slope Diseased  4.4.1  0.40 0.43 -0.10 0.23  0.12 0.13 0.31 0.41  0.16 0.30 -0.69 -0.19  0.76 0.57 0.65 0.64  n 38 3 38 3  Usability Evaluation Goals  The overall usability goals can be stated as 1) to determine the speed at which the users can use the system for screening and determining whether it is adequate for our purposes, and 2) to determine the ease of use (or usability) of the system. Determining the functionality of the system is an aim of the patient study, described in Section 4.5. Determining usability can be further broken down into smaller subgoals. We chose to focus on specific aspects of system usability, which are: • Learnability - How easy it is to learn the system. • Satisfaction - How satisfied users are using the system. • Usability evaluation to uncover problems that have yet not been identified. A usability evaluation protocol was prepared as part of the MITACS funded project ‘Usability of vessel compressibility characterization system using ultrasound imaging.’ This protocol outlines the usability goals along with specific tasks and measurements needed for the evaluation. The main points are presented here. Volunteers were invited to participate in the user study from the pool of potential users, including ultrasound technicians, ward nurses and medical students. User’s level of expertise with ultrasound and the DVT screening system (if applicable) was recorded. Healthy subjects were recruited from the general population as volunteers for the system users to scan. Subjects with history of DVT were not accepted. All participants were asked to sign the appropriate consent forms in order to participate. User evaluations were approved by the appropriate ethics board (please see DVT Evaluation, UBC Clinical Research Ethics Board, in Appendix E) The user study consisted of two rounds. The first round comprised the learnability, scan time, and user satisfaction evaluations. The second round comprised the memorability, scan time, and user satisfaction evaluations. Error rates were recorded for both rounds, and also included an informal interview. Users took part in the second round no sooner than one month after participating in the first round. Users were considered to be novice users. However, because of the concurrent clinical exam, some users accumulated more experience and were classified as expert users. Written instructions were prepared for each of the tasks in the usability study (described below) and used as the main source of information. Users were not trained to use the system, other than receiving a short demonstration before each session, that lasted no more than five minutes. 115  Learnability Four user tasks for measuring learnability were defined, and the time needed to complete them was measured. A target time was set for each task, and the evaluation proceeded once the user completed the task within this limit. The number of attempts necessary to perform each task within the target times was also recorded. • Tasks: The user was asked to use the DVT screening system to perform the four main tasks which are needed for screening. These were 1) initializing contour detection and tracking, 2) 3-D vessel model construction, 3) vessel compressibility assessment, and 4) vessel flow assessment. An instruction sheet (found in Appendix D.2) was given to each user. • Measurement method: The time taken to complete each task was measured after the user had read the instructions. The measurement began when the user indicated they were ready to start, and ended once the investigator determined that the task had been completed. The user had access to the instruction sheet throughout the procedure. If the user did not complete the task within the predetermined time, they were asked to repeat it. Task time and number of attempts were recorded using a stopwatch and a data collection form. The cumulative time of the evaluation was also recorded. • Evaluation: It is known that the system developer can easily perform all tasks in under 2 minutes for a single vessel. Users had to complete the four tasks (contour detection, model building, compression exam, flow assessment) in one, one, two and two minutes respectively, for a total of 6 minutes in total. Two measures were generated to assess the learnability of the system. The first measure consists of the cumulative completion time (in seconds) divided by 360 (6 minutes), and where smaller numbers indicate better learnability. The second measure takes into account the number of attempts by calculating learnability = i  1 attemptsi · timei  (4.6)  where i represents each of the four tasks, timei is the cumulative time for each task (in seconds), and attemptsi is the number of attempts to complete each task. Here a larger number indicates better learnability. The goal was to obtain a baseline measurement from different groups of users and to assist in the calculation of system memorability. Memorability This aspect refers to how easy it is to remember the system after an extended period. Users were asked to return at least one month after the initial test to repeat the tasks. The difference between the learnability obtained in the two rounds indicates the memorability of the system. • Tasks: The user was asked to use the DVT screening system to perform the four main tasks as described in the learnability experiment. The instructions for this task are in Appendix D.3. However, step-by-step instructions were not given as in the learnability test. 116  • Measurement method: The same as for the learnability experiment. • Evaluation:  The difference between the calculated learnability indices were used as a  measure of memorability. The greater the differences, the more memorable the system is. User Satisfaction The user’s satisfaction with the system performance and interface, as well as their perceived ease of use was measured using questionnaires. For users with significant previous ultrasound experience, we also measured their perceived acceptance of our system compared to the conventional methods. • Tasks: After a user performed the learnability or memorability test and patient scan, the users were asked to answer a questionnaire (Appendix D.4) regarding their satisfaction with the system. • Evaluation: For each question, the average response was calculated. In addition, the maximum and minimum value for each question was recorded. A 5-point scale was used as well as inverted questions (to avoid a ‘check-list’ response from users). Depending on the question, the satisfaction level with the system will be ‘Acceptable’ if the average response is less(greater) than or equal to 2.5(3.5) and less than 20% of users answer with the worst option, ‘Excellent’ if no users answer with the worst option, and ‘Unacceptable’ otherwise. Scan Time The time needed for a user to scan a healthy volunteer using the DVT screening system was measured. In addition, the allotted time for conventional scans as well as the actual scan time was recorded and compared to scan time using the screening system. In addition to data collected during the usability study, scan time obtained during the hospital study (described in Section 4.5) was also used to determine how long it takes to scan a patient using the DVT screening system. • Task: The user was instructed to scan a patient’s vessel as described in the scan protocol in Appendix D.1. This required the construction of several 3-D vessel models using the DVT screening system, as well as performing compression examinations and flow assessments. • Measurement method: Scan time was recorded using a stopwatch and a data collection form. When using the DVT screening system, scan time for each vessel was automatically logged by the system software. • Evaluation: Given the known goals for the DVT system and known estimates for conventional ultrasound screenings, total scan times were classified as follows: less than 5 minutes as ‘Perfect’, more than five and less than 8 as ‘Excellent’, more than 8 minutes and less than 15 as ‘Good’, more than 15 and less than 20 as ‘Adequate’ and more than 20 minutes as ‘Inadequate’. These classifications were subject to review based on the data collected from the conventional scans.  117  The measurement of total scan time included any time needed to configure the system (as we are evaluating the interface). The measurement did not include ‘meet and greet’ time or the time needed to position a patient or answer questions, as it was assumed that these would be similar in both types of scans. In the case of the conventional examinations, the time between the moment the scan is finished and the moment the results of the scan are OK’d by a radiologist were measured separately. The DVT screening system performs an automatic DVT measure calculation, similar to the verification by radiologist, and it is of interest to identify the amount of extra time needed in order to finalize the diagnosis. Informal Interview In addition to the user satisfaction questionnaire, the users were asked about any specific likes or dislikes about the system, as a continuation to item 5 in the questionnaire (‘Please leave any additional comments you may have below:’). The reviewer took short-hand notes about possible aspects of the interface, and some of the system features that were not tested. Error Rate During the entire usability evaluation, the number of errors (both user and software) were recorded as an additional usability performance measure. • Tasks: Learnability, memorability and scan time evaluation. • Measurement method: Using a data collection form the number of errors the user performs will be recorded, including a short-hand description. In addition, any software errors that occur will be recorded. • Evaluation: The user error rate for the system was defined as ‘Acceptable’ if the user had less than 6 errors, ‘Excellent’ if the user had 2 or less errors, and ‘Unacceptable’ otherwise. The software error rate for the system was defined as ‘Unacceptable’ if any software errors occur, and ‘Acceptable’ otherwise. The user errors that were recorded included the user clicking and incorrect button, or the user performing a step out of sync with the program (e.g. by possibly forgetting to click on a button). A software error is any incorrect system behavior - these will be classified as catastrophic (system crash) or other (the system could recover).  4.4.2  Results of Usability Evaluation  Seven users were recruited for the usability study. Scanning sessions took place from December 13, 2007 to March 31, 2008 (round one from December 13, 2007 to February 22, 2008 and round two from January 21, 2008 to March 31, 2008). Two ultrasound technicians, two nurse practitioners and three third year medical students volunteered. 118  Table 4.25: Learnability Indices. Index L1 L2 L3 L4  Description Normalized sum using total time for each task Index using eq. (4.6) using number of attempts and total time for each task Normalized sum using cumulative completion time for each task Index using eq. (4.6) using number of attempts and cumulative completion time for each task  The usability evaluation was performed in a cubicle with a hospital bed in the Robotics and Control Laboratory, Department of Electrical and Computer Engineering, University of British Columbia. Volunteers and users were scheduled for each scan, and volunteers were scanned on several occasions by different users. The DVT screening system based on the Ultrasonix Sonix ultrasound machine was used for scanning. The same software was used for all scans in both rounds. All users and scanned volunteers read and signed the appropriate consent forms (Appendix D.5 and D.6, respectively). A script was generated in order to keep each scan as repeatable as possible. All data was collected on data collection forms generated for that purpose. All users were compensated with a $50 gift card and volunteers were compensated $10 per sitting. Ultrasound experience and ease of interpretation was determined from the questionnaire, questions 4.a - 4.c (questions 18- 19, see Table 4.27), and only two people (ultrasound technicians) had any significant experience (5 and 17 years). Learnability and Memorability The two different learnability indices were calculated using two different sets of data, for a total of four learnability indices L1, L2, L3 and L4. These are described in Table 4.25. Here ‘total time for each task’ refers to the time needed by the user to complete each task from beginning to end, and including time needed to read instructions, modify settings or adding ultrasound gel to the volunteer, the time in between attempts or repeating the task (if so instructed), among other possibilities. ‘Cumulative completion time’ refers only to the time spent on performing the task; if more than one attempts was performed, the time for each attempt is added. In general, smaller numbers are desired for L1 and L3, while larger numbers are better for L2 and L4. Resulting learnability index values for both rounds are shown in Table 4.26 and Figure 4.19. The Table includes mean, standard deviation, minimum, and maximum values. Because of missing data, some indices could not be calculated. Therefore, the number of indices used is also displayed (n). Figure 4.20 shows resulting learnability indices by user group. Overall, there is a significant difference between all learnability indices in round one and round two, at a significance level of 10% (two-sample t-test). At a 5% significance level, all indices except for L3 (p-value = 0.05) have significantly different values in round one and two. However, strong conclusions cannot be drawn because of the small sample number. Even so, it is clear that the 119  Figure 4.19: Box plots of learnability indices L1, L2, L3 and L4, calculated from user data in rounds 1 and 2. The boxes have lines at the lower quartile, median, and upper quartile values. The whiskers are lines extending from each end of the boxes to show the extent of the rest of the data. Outliers are data with values beyond the ends of the whiskers. learnability indices generally improve from the first to the second round. The ‘best’ learnability indices for all four indices were obtained from medical students, while the ‘worst’ learnability indices were obtained from the medical students and nurses groups. This only indicates that the group of medical students found it easiest (and at times, the most difficult) to learn the system. In addition, both of these groups showed a large variation, e.g. L3 and L4 in Figure 4.20, in contrast to the group of ultrasound technicians, where the variation in index values is much less. No significant difference (at 10% or 5% levels) was found between the number of attempts at a task between the first and second rounds, indicating that any improvement in the learnability indices was likely due to a faster task completion time rather than a significant decrease in the number of attempts. In terms of memorability, the differences between the learnability indices from the first to second round indicate that in general it is easy to remember how to use the system. When analyzing the values in both rounds, either overall or in groups, there is a clear trend toward better learnability values. This indicates that the system is in fact memorable.  120  Figure 4.20: Box plots of learnability indices L1, L2, L3 and L4, calculated for each group (‘NP’ = nurses, ‘US’ = US technicians, ‘MD’ = medical students) from data in rounds 1 and 2. See Figure 4.19 for box descriptions.  Table 4.26: Learnability Indices - All Users. Index Round Mean Std. Dev. Min. Max.  n  L1  1 2  5.86 2.50  1.11 1.01  4.67 1.33  7.67 3.50  6 4  L2  1 2  0.09 0.33  0.03 0.14  0.06 0.17  0.14 0.50  6 4  L3  1 2  3.30 1.85  1.31 1.06  1.94 0.74  5.75 3.81  6 7  L4  1 2  0.40 1.02  0.26 0.49  0.08 0.16  0.75 1.67  6 7  Scan Time After performing the learnability or memorability tasks, the users were asked to scan the volunteer’s vessels as outlined by the scan protocol. This consists of performing the four predetermined tasks (contour tracking, model building, compression exam, and flow assessment) for each vessel. Because of time limitations, users only scanned vessel sections, and not the complete venous system as would be done on a true patient. However, scan time was recorded for all subjects in the hospital study, from which an additional estimate of scan time is obtained. This is detailed in Section 4.5. 121  Figure 4.21: Box plots of scan time for conventional ultrasound compression examinations, when scanning one or two limbs. Does not include time needed for review by radiologist. See Figure 4.19 for box descriptions. Measurements of scan time for individual vessels, however, are available. In round one, from a total of six vessel scans, a mean scan time of 5.45 minutes (5 minutes, 27 seconds) was obtained overall. In round two, from a total of ten vessel scans, a mean scan time of 3.74 minutes (3 minutes, 44 seconds) was obtained overall. Even though scan time seemed to decrease from the first to the second round, there was no significant difference between the two groups of measurements (twosided t-test, p-value = 0.17). The ultrasound technician and medical student groups showed a decrease of about 30 - 40% in scan time between rounds one and two, while the nurse group remained unchanged. The group consisting of medical students showed scan times with the least standard variation in both rounds (a 1.5 minutes in the first round and about a minute in the second round), while the group of ultrasound technicians showed the largest standard variation (about 4 minutes). In addition, both medical students and ultrasound technicians had the fastest scans in both rounds (2 minutes per vessel in round two). The scan time values in round two approach what was expected for a single vessel, about 2 minutes, in the best case scenario. Maximum scan times were about 8.5 minutes per vessel in round 2 indicating that there is still variation in the data. Mean values suggest a complete examination in about 15 minutes using the DVT screening system is possible. Finally, to determine the time taken for conventional compression examinations, measurements were taken at UBC Hospital between November 6, 2007 and February 4, 2008. A separate entry was made for the additional time needed for a radiologist to review each scan and give a final diagnosis. A total of 32 scans were timed, of which 26 were for single leg examinations and 6 were for examinations involving two limbs. The time was recorded by the sonographer performing the examination. Figure 4.21 shows the scan times obtained, grouped by the number of limbs scanned. The mean compression ultrasound scan time when scanning one limb was 13.85 minutes (13 minutes, 51 seconds), while the time when scanning two limbs was 22.33 minutes (22 minutes 20 seconds). It was found that there was a significant difference between scan times for one limb 122  Table 4.27: Questionnaire Statements. Num. 1 2 3  4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19  Statement When learning how to use the DVT screening system I found that... ...the instruction provided was enough. ...it was easy to learn how to operate the system. I feel that I could properly use the DVT screening system again with little or no additional instruction. When using the DVT screening system I found that... ...the tasks were easy to understand. ...the tasks were easy to perform. ...the final objective was clear. ...the displayed information was useful. ...there was too much information displayed. ...there was information that was missing. ...controlling the system using the interface was easy. ...the interface made the tasks more difficult to perform. ...using the vessel tracking was easy. ...making a 3-D vessel model was easy. ...performing a compression examination of a complete vessel was easy. Overall, I think that the DVT screening system is easy to use. It is easy for me to interpret ultrasound images. It is easy for me to interpret ultrasound images of vessels. Years of ultrasound scanning experience (0 for no experience): Please leave any comments you may have below:  and scan times for two limbs (two-sample t-test, p-value = 0.0024), but there was no significant difference in review times (p-value = 0.90). The pooled review time was 7.16 minutes (7 minutes, 10 seconds) per patient. This is consistent with the longer reported scan times when performing bilateral examinations. User Satisfaction At the end of each round, user answered a questionnaire (Appendix D.4) regarding their impressions with the DVT screening system. Nineteen questions were included, which are shown here for reference in Table 4.27 with a corresponding question number. For the first seventeen questions, users were asked to answer by circling a number from 1 to 5 for each statement that corresponded to their response, where 1 indicates that they strongly agree, 3 indicates that they neither agree nor disagree and 5 indicates that they strongly disagree. The final two questions were open. Overall user responses are presented in Table 4.28 and Figure 4.22. As outline above, the satisfaction level for each question is ‘Acceptable’ if the average response is less than or equal to 2.5 (if the desired response is 1) or greater than or equal to 3.5 (if the desired response is 5) and in addition less than 20% of users answer with the worst option. If in addition to this no users 123  Table 4.28: Overall Questionnaire Responses. See Table 4.27 for key to question numbers. Round 1 Round 2 Num Mean Median Min. Max Satisfaction Mean Median Min. Max Satisfaction Level Level 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17  2.0 2.1 2.7 1.4 2.1 1.6 1.9 3.9 4.3 2.4 3.6 2.3 2.1 2.4 2.3 2.4 2.4  2 2 2 1 2 2 2 4 4 3 4 2 2 2 2 2 2  1 1 2 1 1 1 1 3 4 1 2 1 1 1 1 1 1  4 4 5 2 3 2 4 4 5 4 4 4 4 4 4 4 4  Excellent Excellent Unacceptable Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent  1.5 2.0 2.0 1.9 1.7 1.4 2.1 3.6 3.5 2.4 3.4 2.4 2.3 2.4 2.0 2.3 2.3  1.5 2 2 2 2 1 2 4 4 2 4 3 2 2 2 2 2  1 1 1 1 1 1 1 2 2 1 2 1 1 2 1 1 1  2 3 4 4 2 2 3 4 4 4 4 4 4 4 3 4 4  Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Excellent Unacceptable Excellent Excellent Excellent Excellent Excellent Excellent  124  Figure 4.22: Box plots of user responses to questionnaire from rounds 1 and 2. See Table 4.27 for key to question numbers. answer with the worst option (e.g. answer 5 when the desired response is 1) then the satisfaction level is ‘Excellent.’ Otherwise, the user satisfaction will is ‘Unacceptable.’ In addition, user comments (question 19) for each round are presented in Table 4.29. Overall, user satisfaction with the system is good, except for one question in each round. Otherwise, there are small changes from round 1 to round 2, and these are mostly improvements in the user’s perception and satisfaction with the system. These trends can also be observed when grouping users based on their profession (data not shown). Finally, a recurring comment is that the ultrasound probe and shells are heavy. Error Rates Throughout the user evaluation, user errors were recorded. At times, users were asked to use the ‘talk out loud’ approach as they were performing the examination, in order to gain insight into any difficulties that they may have been experiencing. Any software errors that occurred were also recorded. Tables 4.30 and 4.31 show user errors and frequency of errors (n) for both rounds. If a user recovered from an error, it is so indicated at the end of the error description. Table 4.32 shows user comments obtained during the ‘talking out loud’ approach. 125  Table 4.29: Questionnaire Responses - User Comments. Round 1 Will be better next time! It would be helpful to attend a basic inservice on US technique and vessel locations. [W]ritten instructions [were] not easy to follow. Probe too heavy. Probe does get heavy after a while. [I found] it ... easy for me to interpret US images. This is great because I was not able to do so in the past. Sometimes difficult to see vessel on screen because lights on. Changing the ultrasound settings (i.e brightness, contrast) made the tasks much easier for me. Round 2 It is getting easier to follow instructions on my second session. My understanding of instructions and finding the vessel was improved compared to first time. [Include] a diagram of venous anatomy on the leg [to assist during scan]. May be helpful if instructions for compression exam state to press 3 each time you are going to compress the vein. [H]aving the number ‘hot keys’ displayed [on the interface] would be helpful. Controlling the system using the interface was difficult [using the trackball]. If controlling system is on the probe it may be easier? Probe was heavy. Notes were also made of ‘good’ user behavior, such as when users recovered from potential errors on their own, or used some of the advanced features. During the scans in the second round it was observed that at least 2 users remembered the numbered button shortcuts or hotkeys to perform the tasks, at least 3 users remembered how to reinitialize the contour detection and tracking using the saved seed point, and at least one user used the 3-D vessel model to assist in orienting the ultrasound probe in the longitudinal plane during a flow assessment. System errors found are presented in Table 4.33. Potential or actual solutions to these errors are also included in the Table. In addition, several observations were made by the author during these scans. When the user first presses ‘Start’ and prepares to begin a scan, a force calibration dialog box appears and asks the user to place the probe gently on the patient, with the objective of calibrating the force sensor and obtaining a baseline force measurement to be used as ‘zero’ applied force. The implementation however does not display the ultrasound image – instead the user must switch windows to view the Sonix display. Future implementations should include the prompt for calibration and the ultrasound image on the same screen. Another observation is that the method of acquiring data during the colour mode is different to that used when building a model or performing a compression exam. In the latter cases, the 126  Table 4.30: User Errors, Round 1. n  Error  1  Examiner does not hold probe steady when moving pointer to the ‘Find’ button (after having selected seed) Difficulty mapping probe movement (move left, image seems to move right) During model building, user drags probe along leg before contour is on User does not turn off make model button after building model User tries to do compression exam without first pressing corresponding button User tries to do compression exam without first pressing corresponding button - USER RECOVERS User aborted compression exam because seed lost (instead of restarting tracking) User tries to perform compression exam without a seed User tries to perform compression exam without a seed - USER RECOVERS User did not complete compression exam as per instructions; user thought they had finished but there were model locations without data User did not see force bar User gave flow instructions (e.g. flex your ankle) to volunteer patient too slowly; might also indicate flow exam timing is too fast To repeat a flow examination, user clicks on button (which exists the flow examination mode) instead of clicking on the ultrasound image  3 1 2 2 1 1 1 1 1 1 1 1  user must first select a button, then perform the task. In the former case, the user selects a button to switch modes 16 and then the user must select the location on the image where the colour flow is located. During the scans this difference was seen to lead to confusion on how to perform the colour assessment, or more specifically when the data acquisition step begins. User errors in Tables 4.30 and 4.31 may also stem from this. Finally, a repeated error was that users attempted to perform compression examinations without having a valid seed and contour. Currently, the system will not allow a user to perform a compression examination if a 3–D model has not been built. The system automatically stops the compression exam and places a message indicating what has occurred. However, stopping a compression exam in a similar way may lead to increased confusion, as it was noted that users did not pay attention to the status messages.  4.5  Clinical Study of the DVT Screening System  The DVT screening system was evaluated in a clinical setting at the University of British Columbia Hospital. The objective of the study was to determine if the system’s functionality was adequate for detecting DVT, by comparing the results obtained using the system to the results of conventional compression ultrasound examinations. A protocol for the study was presented to and approved by the appropriate ethics review boards (approval forms can be found in Appendix E). The study was performed using the Sonix-based system described in Section 3.5.1, with the 16  The use of modes is actually discouraged [140].  127  Table 4.31: User Errors, Round 2. n  Error  1 1  US image difficult to interpret; difficulty finding vessel User expects tracking to start when they click on image; they forgot to click on the ‘Find Vessel’ button first User tries to build model without first clicking on model button User tries to build model without first clicking on model button - USER RECOVERS User forgets to click model build again when finished, switching it’s state to ‘off’ Compression exam procedure not clear, user is unsure of what to do User forgets to re-click on compression exam button each time they compress User tries to do compression exam without first pressing corresponding button User tries to do compression exam without first pressing corresponding button - USER RECOVERS User tries to perform compression exam without a seed User tries to perform compression exam without a seed - USER RECOVERS User doesn’t look at/interprets toggle state of button and clicks again In the flow assessment, user incorrectly associates blue/red with vein/artery User not clear on how to exit (to press the ‘Review’ button) User clicks on button instead of clicking on the ultrasound image to repeat a flow examination - USER RECOVERS  1 1 1 1 2 1 1 1 1 2 1 1 1  user interface that was evaluated (Section 3.5.3) and then modified based on the usability results (Section 4.4.2). Major differences comprised the addition of a touchscreen (Magic Touch X-Model, Keytec, Inc.), a vessel specific tracking model for DVT (Section 3.2.8), and improved system messages. Patients that had undergone hip replacement surgery or knee replacement surgery were invited to participate. Only those who could provide informed consent (see patient study consent form, Appendix D.7) were scanned. The DVT screening system was used to scan segments of the superficial femoral vein and assess them for compression. Vessel bifurcations were not scanned. When possible, more than one vessel segment per patient was scanned. When possible, calf veins were scanned. The examinations approved by the UBC Clinical Research Ethics Board were carried out by two nurse practitioners, Leah Christoff and Jean Hong, whose names were included in the research protocols. The scanning protocol consisted of an initial scan to build the 3-D model with the detected contours, followed by a second scan to assess the vessel for compression. Compression data was gathered in compression-release cycles. Finally, a flow assessment was performed by measuring mean venous flow values, determined by colour ultrasound, during rest and when the patient flexed their ankle. The two examiners that performed the scans participated in the user study. In addition, several practice scans with healthy volunteers were done in order to train the examiners. The examiners felt that they could use the system appropriately. 128  Table 4.32: User Errors - ‘Talking Out Loud’ Comments. n Comment Round 1 1 1 1 1 3  ‘Purple’ is ambiguous; describe the color in more detail The interpretation of the ultrasound image is difficult The ultrasound cord is heavy / gets in the way Changing the ultrasound image settings made the tasks easier Written compression exam instructions not clear, too technical  Round 2 2 1 1 1 1 1 1 1 1  4.5.1  Instructions are not specific enough, need improvement The flow assessment happens too fast The ultrasound image is ‘backwards’ Changing the 3-D viewpoint (from the default side view to front view) made scanning easier The pop-up dialog with flow assessment instructions gets in the way, and I [user] do not pay attention to them; they should maybe only appear once The trackball is difficult to use The probe cable is heavy A model length indicator would be useful to determine appropriate model length System messages in the control display are ambiguous; it is not clear which is the current message  Results of Hospital Evaluation  The hospital evaluation took place between April 8, 2008 and May 16, 2008, at the University of British Columbia Hospital, part of Vancouver Coastal Health Authority. A total of 11 patients were recruited (5 male, 6 female) with an average age of 69.8 (minimum 56, maximum 77). Patients were classified as having light (1), medium (7), or heavy (3) build. No patients had a known history of DVT, however all patients had either undergone total hip (1) or total knee (10, including 2 bilateral) replacement surgery in the preceding days. Results from the conventional CUS examination were used as the gold standard. Four patients were found positive for deep vein thrombosis, of which one was above-knee DVT in the CFV and three were below-knee DVT (1 POP, 2 PER, PT). Sonographers reported difficulty visualizing the calf veins in 5 cases, including the two patients with positive DVT in the CFV and POP vein. When using the DVT screening system, the nurse practitioners were unable to visualize the deep veins in two cases, including the patient positive for DVT in the CFV. The other case with which the nurse practitioners had difficulty visualizing the veins was negative for DVT on CUS. Vessel models were not made in these cases, nor were the vessels assessed for compression. However, data was saved for offline processing. A total of 18 vessels were scanned using the DVT screening system, of which 11 (7 patients)  129  Error  Table 4.33: System Errors. Solution  Contour order in 3-D model is incorrect Sonix image freezes during scan (due to Ultrasonix software) Force offset used during seed tracking too big Force offset ‘pushes’ seed down during release in a compression exam System does not indicate to the examiner when to tell the patient to relax during a flow assessment Flow assessment timing is too fast Sonix does not keep depth/gain/focus settings when switching between flowvenous modes Calf veins too small to track  Vessel tracking errors due to duplicated veins, during compression examination, because tracking latches on to the artery, or poor image quality Ultrasound color flow aliasing generates erroneous flow assessment measures  Do not sort contours during model building None at this time Reduce relative value of force offset Reduce relative value of force offset, and/or have a different force offset during compression and release Added a message at the end of the flow assessment to tell patient to relax, to stop flexing their ankle Make the flow assessment slower Use a custom Sonix preset to save current settings each time user changes modes Generate a system preset to change contour detection and tracking settings based on type of vessel being tracked (thigh vein or calf vein) Modify tracking model so that tracking the artery is not allowed; image quality can be improved only in some cases Train users to identify and correct color flow aliasing  were from left limbs and the remaining 7 (4 patients) were right limbs17 . The mean scan time per vessel was 417.64 sec (6.96 minutes) with a standard deviation of 174.55 sec (2.91 minutes), and a minimum of 174.96 sec (2.92 minutes) and maximum of 698.30 sec (11.64 minutes). Scan time included building a vessel model, performing a compression examination, and performing a flow assessment. As described under the ‘Results’ heading in Section 4.3.3, various vessel compressibility measures were calculated for each scanned vessel, using data assigned to each vessel model. Using the average TAR values for each vessel model, the sensitivity and specificity of the system are 25% and 92.86%, respectively, similar to the values obtained in the pilot study. When using the maximum TAR values for each vessel model, sensitivity increases to 50%, while specificity drops to 85.71%. When using the slope measure, a sensitivity of 100% is obtained, and specificity ranges from 35.71–64.29%, depending on whether the maximum or average slope value for each vessel model is used. In addition, the average slope values for patients with DVT show a significant difference (t-test, p-value = 0.015632) from average slope values for patients without DVT. As before, the 17  Both legs were scanned on a bilateral knee replacement patient.  130  Measures J value  maxTAR 1.3975  Table 4.34: J values for vessel measures. meanTAR maxSlope meanSlope maxSlopeF 1.711  1.4483  1.4577  meanSlopeF  flow  1.3666  1.7368  1.622  slope measure outperforms the TAR measure, suggesting that the slope measure is less sensitive to noisy measurements. Blood flow assessment was performed using the vessel characterization system on 13 vessels (10 patients), of which 4 vessels (3 patients) were identified as positive for DVT by CUS. Of these scans, 3 vessels (3 patients, all negative for DVT by CUS) were excluded because of errors during data acquisition (incorrect cursor placement, probe movement) or patient movement. The mean flow ratio for patients that were negative for DVT was 1.3 (minimum 1.0, maximum 1.694) and the mean flow ratio for patients positive for DVT was 1.009 (minimum 0.948, maximum 1.098). These mean values show a significant difference (t-test, p-value = 0.09691, α = 0.1). For cutoff values ranging from 1.1 to 1.4, the sensitivity and specificity for this measure are 100% and 66.7% respectively. Optimal Feature Calculation The performance of the vessel characterization measures was evaluated based on a classification scheme presented in [124]. Data is collected and assigned to one of two classes based on the presence or absence of DVT, denoted by X 1 and X 2 respectively. A value J is then calculated as J=  Tr(Sm ) Tr(Sw )  (4.7)  where Sm is defined as the covariance of all the measurements X = [X 1 X 2 ] or Sm = E{(X − µo )(X − µo ) }  (4.8)  given that µo is the mean of X , and Sw is the weighted covariance Sw =  N1 N2 S1 + S2 N N  (4.9)  where N1 , N2 are the number of measurements for each class, N = N1 + N2 or the total number of measurements, and S1 and S2 are the covariances of X 1 and X 2 respectively. A value of J = 1 indicates that the specific measure is only as good as chance for differentiating between the two classes. Higher values of J indicate an increasing ability to correctly differentiate whether a measurement belongs to one class or the other. An optimal measure was then calculated using [124] y = (µ2 − µ1 ) Sw−1 X  (4.10)  where µ1 is the mean value of the measurements in class 1 (positive for DVT) and µ2 is the mean 131  Table 4.35: Mean (µ), standard deviation (std. dev.), maximum and minimum for optimal feature, for positive and negative DVT classes. Class µ std. dev. Maximum Minimum Positive Negative  786.46 1065.02  23.61 8.31  812.8 1073.0  767.2 1054.2  value of the measurements in class 2 (negative for DVT), and X is a vector containing all the measurements from both classes. The J value for this optimal feature can then be calculated, and is compared to the previous values to determine whether the optimal, combined feature provides additional information. Vessel scans that had compressibility data (TAR, Slope) and flow data (flow ratio) were considered for the calculation of a combined measure to determine DVT. If a vessel scan did not have at least one of these measures, it was not included in this analysis. These scans were classified as positive (3) or negative (4) for DVT, for a total of 7 scans. For each vessel model, several features were used for this analysis. These were the maximum (maxTAR) and mean (meanTAR) Transverse Area Ratio values for each vessel model, the maximum (maxSlope) and mean (meanSlope) Slope values for each vessel model, the maximum (maxSlopeF) and mean (meanSlopeF) non-normalized Slope values for each vessel model, and the flow index (flow) for the scanned vessel. The non-normalized slope value is calculated in the same manner as the normalized slope, except that force values are not normalized. Table 4.34 shows the resulting J values for the individual measures. A value of 1 indicates that the measure is only as good as chance for differentiating between the two classes. The system sensitivity and specificity reported above were calculated using the maximum and mean TAR and slope values. An optimal measure was then calculated as described above, and a value of J = 80.3799 was obtained. Mean, standard deviations, maximum and minimum values of the resulting optimal feature values for the two classes (positive and negative for DVT) are presented in Table 4.35. It is clear from the results that there is no overlap between positive and negative classes, and there is a very clear separation between patients positive and negative for DVT. Using a leave–one–out approach, the mean and covariance values were calculated using a training subset and then used to determine the classification of the remaining dataset. Using this approach, all datasets were classified, resulting in a sensitivity of 100% and specificity of 75%.  132  Chapter 5  Discussion and Conclusions In previous Chapters, the development and evaluation of a system for vessel characterization, specifically for detecting deep vein thrombosis, was presented. The implementation of this system draws on knowledge from diverse areas such as medical imaging, image processing, robotics, and human-computer interaction, among others. The evaluation of the system and sub-systems include data and experiments conducted using simulations, in a laboratory setting, and clinical evaluations. This Chapter presents a discussion of the material presented in this dissertation, and concludes with an overview of system for vessel characterization project and summary of the contributions in Section 5.4. The pivotal sub-systems and components, which are the vessel contour detection, vessel tracking, 3–D model building, validation of the proposed measures for vessel characterization, and the usability of the system interface, were extensively evaluated and are presented below.  5.1  Contour Detection  For vessel contour detection, a formulation for using non-linear models, parameter estimation and extended Kalman filters for contour detection in images has been presented. These models, which use a priori information about the general shape of the feature to be detected to estimate a contour are useful for semi-automated or automated segmentation of ultrasound images, as well as in other images that may have noise or edges that are not clearly delineated. Additionally, an approximate value of the feature area can be directly and quickly calculated. An elliptical model was used to approximate the vessel contour. An ellipse can represent a vessel cross-section under varying degrees of compression, as is observed during a CUS exam for DVT. Using only a few parameters, a smooth curve that approximates the vessel lumen can be generated and an approximate transverse vessel area can be quickly calculated. The representation in many cases will be quite accurate, while significant errors will occur in others. However, the use of an ellipse as a model for the contour detection provides a good search region for the edge detection, and is not used to parameterize the edge. The 3–parameter ellipse model, when using the output approach (Section 3.1.2) was shown to converge to the correct parameter values (Section 4.1.1) in about 95% of attempts, when using strict benchmarks (parameter errors less than 1%). Initial conditions play an important role in the parameter estimation, and are mostly the cause of the over 5% of cases of inadequate performance. When initial conditions are adequately selected, that is we approximately know the size of the feature we are looking for, results are improved. When evaluating the 3–parameter contour detection and parameter estimation algorithm on 133  simulated and patient images, it was shown that the algorithm can identify ellipse parameters with mean errors ranging from ∼2% to ∼13%, and can detect contours that correspond to expert tracings. The estimated contour area also correlates to the true transverse area as determined by the experts, confirming that the quick and parametric calculation of area of elliptical features is accurate. Even when features are slightly non-elliptical, area calculations reflect true values when using this algorithm, as seen by the moderately low mean area values in Table 4.11. This may be due to the over- and underestimation of the ellipse parameters at different locations on the contour, which result in an ellipse with similar area to that of the vessel. Furthermore, mean errors between the detected contour and expert tracing are low. Results across experts also are consistent. This contour detection algorithm performs well when compared to other methods. When comparing the detected contour to expert tracings, we have obtained a mean error of -1.0 pixels with a standard deviation of 2.16 pixels on simulated images, and a mean error of 1.7 pixels with a standard deviation of 5.17 pixels on patient images. For example, [92] reports a mean error of 2.98 pixels with a standard deviation of 3.79 pixels, detecting features ranging from 6.7-10.7 mm and scaling factors of 0.067 mm/pix, very similar to those in our DVT screening system. While our method is more variable, the trade-off is greatly decreased processing time. Slightly better results are obtained as the validation threshold defined in (3.1) is lowered. While there is some variability in the results, by observing the standard deviation in Table 4.2, this should be reduced when used in conjunction with a tracking algorithm. Also, some error may have been introduced into the ellipse parameter measurements because of the possible ’rotation’ of the results if a and b are switched and φ rotated by ±90◦ . Additionally, the result is expressed as a percentage of the true value, yet it can be seen in Table 4.3 that the standard deviation of both a and b are quite similar, and are not more than 6 pixels (∼10% of the maximum feature dimension). Moreover, our contour detection algorithm has been proven to perform reliably over a range of SNRs, as consistent results were obtained for tissue SNRs between 40 to 12 dB. Indeed, statistically significant differences of all error measures are not found until the magnitude of the noise approaches the magnitude of the signal (SNR = 0dB). Typical expected SNRs range from 60 to 40 dB for tissue in B-mode images. While all results obtained from images with attenuation are significantly different from those obtained from images without attenuation, the differences between mean errors are very small (< 4%), and in the case of the ellipse angle φ the error is reduced. Additionally, the spread of all errors are also reduced. This variation in parameter estimation and the associated transverse area values, is deemed acceptable for the DVT screening system. The large variation in φ can be explained by the two circular images used (e = 0), as in these cases φ can take any value. When removing these images from the dataset, the mean value of φ is -3.26◦ with standard deviation of 10.78◦ , when using max erRM S = 6 and n = 546. This variation can further be reduced by taking the mean value of φ over all k, resulting in a mean value of -2.16◦ and standard deviation of 5.13◦ (max erRM S = 6, n = 738). In the DVT screening system, results from the contour detection from the previous frame are used to initialize the current frame. While this was not included in the current validation, it has been observed that the contour detection is stable over successive frames obtained at the same 134  location. Future work could include an evaluation of the reliability of the contour detection with data from previous frames compared to standard initial conditions. The use of an ellipse model was further explored with the development of a 5–parameter ellipse model (Section 3.1.3). This approach includes the estimation of the ellipse centre in addition to the ellipse axes and angle with respect to the image reference axis. While both this approach and that described above are based on ellipse models, the extension of the contour detection from three to five parameters is not trivial. An appropriate output function must be found that relates the ellipse parameters to the edge measurement provided by the ‘Star’ algorithm framework. Results for parameter convergence for the 5–parameter ellipse model are similar to those obtained with the 3–parameter version. In addition, the 5–parameter contour (5PAR) detection algorithm was evaluated using simulated and patient data and the results were compared to the 3–parameter (3PAR) implementation. In these experiments, the seed points were chosen to accentuate the strengths of the 5–parameter model. When all seeds are at the centre of the vessel, the results from the 3– and 5–parameter models are very similar. Indeed, when the ellipse centre xe equals the coordinate origin xc in Figure 3.3, the 5–parameter model can reduced to the 3–parameter model with a single ρ. The data clearly show that using 5PAR results in good parameter estimation and smaller error variations when compared to the 3PAR algorithm, as seen by the generally smaller σ in the Tables in Section 4.1.2. While the ellipse area (AE) from 5PAR overestimates the expert tracings in Tables 4.16 and 4.17, correspondence is good as seen by AOC and AOE. In addition, an estimate of the ellipse centre is obtained only when using a 5–parameter model. The advantages of using a 5–parameter model over one that does not include the ellipse centre is that the initial seed point is no longer required to lie very close to the centre of the vessel or the feature to be detected. Errors introduced by factors such as imprecise user input or inadequate seed point tracking can be reduced using a 5–parameter model. With respect to parameter convergence, a method of re-ordering the search radii to mimic a larger angular step was presented that reduces the computation required to obtain a parameter estimation using a ‘Star’ algorithm. This approach was implemented using the 5–parameter ellipse model (5PAR-l) and compared to a single iteration of the same model (5PAR-s) using a small angular step. It is clear from Table 4.12 that parameter convergence is faster when using a large angular step. This can also be observed when comparing estimated parameters using (5PAR-s) and (5PAR-l) in Table 4.14. Indeed, parameter estimation using a large step and a single radial traversal is comparable to that obtained when using the small step, multiple traversal approach. In addition, there is less variation in parameter errors when using a large angular step. However, the detected contour is not as accurate with a large step as seen from larger RMSEr and lower AOC values. This is due to errors encountered at the start of the estimation procedure. With multiple traversals, large errors due to initial conditions are not seen since each contour point is estimated several times. With a single traversal, contour points are calculated only once, and any large initial errors will be made evident. For applications where only an estimate of the general shape and transverse area is necessary, 135  using the larger step size and fewer iterations provides a good option. A combined estimation first using large and then small step sizes can also be used in cases where both accurate parameter estimation and contour detection are required. While the reported execution time indicates the 5–parameter algorithms take longer than the 3– parameter model, real-time implementation (20 Hz) does not pose significant problems. Moreover, our mean segmentation time of an optimized version of the 3–parameter model takes ∼23 ms, which is much lower than other segmentation methods (for example, [92] reports 0.8 s for 128×128 images, on a Pentium II 300 MHz computer). The 5–parameter model has a similar execution time. Precise measurements for the 5–parameter model have not been obtained. One assumption with the DVT screening system is that vessels will be generally dark. If a thrombus is in the vessel, this might not be the case. While chronic clots may be more echogenic, younger clots are more likely to have echogenicity similar to that of blood [42]. The manner in which clots are located using the CUS method also supports this observation. The system in which the contour detection and tracking is implemented is aimed at screening patients at risk of DVT for new thrombi, not chronic thrombi, as immediate medical treatment is important in avoiding DVT complications [4, 42, 96]. A final extension of the ellipse model was to use an extended Kalman filter for parameter estimation of a 3–D object using an ellipsoid model. As when going from three to five parameters in 2–D, extending the approach to use an ellipsoid model was not trivial. The resulting algorithm can estimate almost all parameters that uniquely define an ellipsoid, including the 3 semi–axes, the ellipsoid centre in 3–D, and the rotation of the ellipsoid about the z axis (rotation about the x and y is assumed zero). In particular, this configuration may be useful for segmentation of the prostate in ultrasound images obtained transrectally. Parameter estimation using this ellipsoid model is seen to converge, with results similar to those obtained with the previous models. While these are only preliminary results, they indicate that the approach developed for parameter estimation in 2–D images may well be used (with appropriate modifications) in 3–D as well. Several additional features have been implemented in the contour detection, as described in Sections 3.1.5 through 3.1.7, which include the use of an angle-dependent measurement covariance, multiple measurements indicating the location of the edge, and specifying the initial radius angle to correspond with the y direction in the image. Of these, perhaps the most advantageous for contour detection in the real-time system is the use of multiple measurements, specifically the use of a radial similarity measurement. By calculating the similarity of the brightness values along the corresponding radii in two consecutive images, a consistent contour detection from frame to frame is obtained. Previous versions of the contour detection that did not include this feature were seen to provide contours that could change from one image to the next, making the detection appear to ‘twitch.’ The use of a radial similarity measurement allows for a degree of propagation of information across image frames, making the contour detection more consistent.  136  5.2  Vessel Tracking  Various tracking models and combinations of these have been explored for determining the seed location in a series of images in real-time. The models that were used combined known (possibly noisy) data, such as the location and force sensor measurements, the location of a 3–D vessel model, and image brightness, with assumptions on how the seed point or vessel centre ‘moves’ in the image frame. The Kalman filter provides an efficient manner in which to combine this information and generate a single estimate. In terms of the model, the most important component is the inclusion of a delayed measurement. In many cases during tracking, there will be no reliable measurement available other than the location of the seed point at the previous step. Therefore an approach that can use this data to estimate the current seed location is highly beneficial. In a sense, all the additional measurements and the models with increased complexity only slightly modify the estimate under normal tracking. However, there are many situations in which vessel tracking will become a difficult task. Poor image quality, fast movements and changing anatomical structures pose challenges for the tracking. It is in these cases where the additional measurements and complexity are greatly beneficial. The use of the sub-sampled sum of absolutes (SSA) is equally important for adequate tracking. While typically template matching has been used to track a feature over successive frames, we have modified this concept for determining the location of minimum brightness. SAD correlation has previously been identified as a good method for determining correlation between ultrasound image data and a mask [54]. Our method has addressed the issue of high processing time, and we show that the use of a similar approach, SSA, in conjunction with the tracking greatly decreases execution time, without any noticeable decrease in tracking accuracy. We have also shown that the use of SSA provides an improvement over previous tracking systems that use Kalman filters [1]. Because of the developed application, a DVT screening system, a specific tracking problem was identified. During a compression and release cycle, ideally the examined vein should disappear. In terms of tracking and contour, this poses a significant problem since we cannot detect a contour (and therefore track the seed point) if it is not there. Quick recovery from the loss of a seed point and contour during a compression exam is therefore very important. This in part also dictated the manner in which data was collected for vessel characterization during compression (Section 3.4.1). A combination of seed persistence, using the 3–D vessel model as measurement, and using previous compression data as measurement help to insure a quick seed point recovery. The main drawback of the tracking implementations is sensitivity to location measurements. Firstly, location sensor noise may adversely affect tracking accuracy. Secondly, because of the distance from the sensor to the actual image plane location, small orientation errors may be translated into large errors. Even so, preliminary results show that the inclusion of location measurements greatly enhances the seed tracking capability of the system, and any deterioration because of measurement noise is greatly outweighed by the ability of tracking fast movements. The authors are aware that the presence of metals has to be accounted for through calibration of the location sensor (the ultrasound transducer shell is made of aluminum). The accuracy was tested using a qualitative evaluation on known phantoms and was found to be sufficient for it not 137  to interfere with our application. Patient movement currently cannot be accounted for by the system. This, as well as other sources such as arterial pulse, are other possible sources of error. Several different approaches were used to detect vessel bifurcations. In general, those approaches that only rely on pixel intensities of B-mode images were not found to be useful for detecting these specific anatomical features. The main reasons for this are most likely related to two main factors. Firstly, while vessels in ultrasound images can be generalized as having a dark interior surrounded by a bright region, other structures in the image may also fit this description. Second, because of the immense complexity of the human body, a bifurcating vessel at the groin or upper thigh may produce a radically different image than a vessel in the popliteal fossa, not to mention at other locations such as the neck. Different surrounding tissues such as fat, muscle, tendons and other vessels contribute to these differences in the generated images. Using individual 2-D slices to detect these features does not seem like a promising approach. However, vessel bifurcations were detected when using additional data. By the inclusion of the information obtained using color flow imaging, it was shown that bifurcations can be detected. This method also has potential for automatic seed initialization based on the identification of flow regions within an ultrasound image. The main drawback to this approach is the limit the ultrasound acquisition imposes on the real-time operation of the system. Because of the manner in which blood flow using ultrasound is calculated, the frame rate of the ultrasound system drops noticeably. The size of the color region of interest also impacts the frame rate. These lower frame rates of 10-15 Hz or less start to make vessel tracking unfeasible.  5.3  DVT Screening System  Several reasons have been stated for the importance of the development of a DVT screening system. It is known that DVT is a common disease that can lead to possibly fatal pulmonary embolism and other complications. Current diagnosis of DVT relies on compression ultrasound examinations, which have been proven more effective than diagnosis from signs or symptoms only. However, diagnosis using ultrasound may take a long time (at times in excess of 40 minutes per limb) and may be operator dependent. Because of limited resources, many health care institutions opt for prescribing anticoagulant medication to high risk patients (e.g., those who have recently undergone hip- or knee-replacement surgery) instead of scanning every single patient. This however can lead to the over-medication of some patients. If a system could provide a fast, user-independent scanning option for DVT, with a high sensitivity, it could improve the manner in which DVT is managed. The measurement-based DVT screening system presented here mimics a conventional CUS examination and therefore comprises several subsystems that perform tasks that a human examiner would traditionally perform. Additionally, a sensorized probe has been developed to permit the acquisition of necessary data, and measures that indicate DVT are calculated to determine the possibility of DVT of an examined venous segment. 138  The use of the additional force and location data provide the necessary information to perform an adequate and complete compression examination. The sensorized probe allows this additional data to be readily gathered, with minimal changes to existing hardware, and without sacrificing the portability of modern ultrasound systems. The DVT system displays the compressibility and general shape of vessels, as well as the relative location of a thrombus, if found, and archives this data for future reference. The vessel model can be used to provide a spatial reference to the examiner and for displaying the results of the examination. Screening for DVT has been suggested by many authors (see pages 17 and 18 in Section 2.2.2). The use of a system as ours could be used in many such situations. A multicentre study cited in [53] reported that screening symptomatic ambulatory patients proved to be a cost-effective solution, where distal DVT occurred in only 0.5% of these patients. The authors state however that a complete examination of the entire proximal venous system should remain the usual procedure in most patients. Another feasible application is screening based on clinical and biochemical parameters, by restricting the test to patients with a high clinical probability score and/or a positive D-dimer test. A strategy such as this proved to be safe and feasible in an emergency department setting [53]. Others [93] suggest that the sequential use of a rapid quantitative D-dimer test, clinical score and CUS appears to be safe and the most cost-effective diagnostic work-up of DVT. Screening with a system such as ours may also be feasible. With respect to the developed measures, the ROC curves for the TAR and slope criteria from Section 4.3.1 provide a good indication of the usefulness of the proposed measures. For both measures, the curves lie to the left of the diagonal on the ROC curves, indicating that these measures in fact do provide useful information [103]. The accuracy of the estimated vessel cross-sectional area will clearly affect the accuracy of the compressibility measures, and this is reflected in the low sensitivity of 60% reported for the slope criterion in this experiment. It is believed that the lower sensitivity for the slope criterion is due to the presence of outliers created from the inability to detect zero area, and clearly indicates that this initial implementation did not match its potential. While experiments on simulated and patient images have shown that mean area errors range from about 5 to 15%, correspondence with expert tracings is quite high, on the order of 90 to 96% for the 5–parameter model on patient images (see Table 4.17, page 103), and areas tend to be over-estimated as seen by generally larger false positive (FP) than false negative (FN) values. This over-estimation of vessel area will in turn generate compressibility criteria values that would ideally correspond to diseased vessels, and possibly lower the system specificity, or the ability of the system to reject non-diseased cases. However, system sensitivity should not be significantly affected. These area estimation errors are therefore acceptable for this application, since false positives for DVT in a screening system would require further testing, but false negatives could be potentially fatal. While we would ideally desire good sensitivity and specificity, good sensitivity for a screening exam is a good alternative. By defining a protocol that requires an in depth examination of all positive results obtained from a screening system with 100% sensitivity (as with the slope criterion) and adequate specificity, we can all but insure that all true positives are detected. 139  5.3.1  Usability  The results obtained during the usability evaluation are encouraging. The measures for learnability mostly set a baseline for the system, but they also seem to indicate that the initial estimates of the time needed to learn how to use the basics of the system were appropriate. However, expert and proficient use would still require significant training and practice. The change in the learnability indices from round one to round two, as well as the results from the questionnaire and the written comments also support this conclusion. Two approaches were taken to calculate two different learnability indices, resulting in four indices, as described in Section 4.4.2. The first approach involved calculating each index using the total time taken during each task, including any time required for setup, reading instructions, etc. (L1, L2). The second approach involved using only the time spent on the task itself (L3, L4), or the cumulative task time. It is clear from the results that the second approach, using cumulative times, generates a broader range of index values. Part of the reason for this may be missing data, but also the fact the a cumulative measurement more accurately reflects the amount of time spent on a task compared to the total time. Small variations between users will most likely show up clearly when calculating the indices using the cumulative time. The number of attempts were treated equally in both approaches. The memorability of the system is closely linked to the learnability, in that the former is calculated from the change of the latter. Again in this case, results are quite encouraging. There are significant difference between the learnability index values in round one and those in round two, indicating that the use of the system is easy to remember. Overall, and even for each subgroup, the mean learnability value improved in round two. Scan times correspond to expected values, especially in the second round. A decrease in scan time indicates that some experience with the system is beneficial. However some variation is still seen, and the longest scan times for a single vessel are unacceptable. Some of the factors possibly affecting scan time include the ultrasound image quality, which will have repercussions on contour detection, tracking and location of the vessels by the user, and ultrasound image interpretation and probe manipulation by the user. This second point will directly affect the speed with which a user can locate a vessel using the ultrasound probe. Users appeared to be mostly satisfied with the system. The only two responses that did not have a satisfactory response were the indicating that the user felt that they could confidently use the system again with little or no training (question 3, in round 1), and that the interface made the task more difficult to perform (question 11, round 2). With respect to round 1 and question 3, ‘I feel that I could properly use the DVT screening system again with little or no additional instruction,’ only one user responded with the worst answer (5, Strongly Disagree) which is less than 20% but the average was greater than 2.5. This indicates that one user did not feel that they could use this system after an hour’s worth of practice. However, data in Figure 4.22 indicates that all other users responded with either 2 or 3, and the median is closer to 2. However, it is not clear whether the source of this lack of confidence lies in the interface and the system, or in some other factor such as familiarity with technology and 140  computers in general. There is however an improvement in round 2, indicating that more training and practice with the system gives the user more confidence to assert that they could use the system with little or no assistance. Finally, it should be noted that even though there was an improvement from round 1 to round 2, the user still disagrees that they could properly use the system again. This indicates that possibly 80 – 90% of users will need only a couple of hours of training and practice, will the remaining users may require more. In round 2, with respect to question 11, ‘When using the DVT screening system I found that the interface made the tasks more difficult to perform,’ user responses in round 1 seem to mostly indicate that user do not feel the interface is interfering, while responses from round 2 show more of a spread. In addition, comments from question 19 and those from the informal interview  18  indicate that as users gain experience with the system, the systems limitations become apparent. In this specific case, the need to use a pointing device (trackball) in order to control the system shows up as a limitation. The final clinical system was modified using a touchscreen overlaid on the ultrasound machine monitor to address this issue. However, this was not extensively tested in the user study because all users used the exact same system in both rounds. An interesting change is observed between round 1 and 2 with respect to questions 8 and 9, which ask whether there was too much information displayed and whether there was information that was missing, respectively. In round 1 users almost unanimously disagree in both cases, but in round 2 even though users still mostly disagree, some start to feel that there is either information missing, or too much information displayed. This again is likely due to users having more experience and feeling more comfortable with the system. Based on the users’ comments, their responses about missing information may refer to such things as including a diagram of the venous system to use as a road map during a scan, or to include ‘hotkeys’ on the button faces19 . Also, one user commented during the informal interview that they were not sure how long a vessel model should be, suggesting that an length indicator, which tells the user when a model is of adequate size, could be helpful. There is no clear answer regarding users’ suggestion that there is too much information displayed. User errors seem to be due mostly to the expectation of what the system does, and what the user is supposed to do. This can also be attributed to a lack of familiarity with the system and protocol, but usability experts point out that there should be no reason that a system tell a user exactly what to do. In this case, our expectation that the user would go through the instructions and remember a series of steps that included selecting several buttons was incorrect. In a sense, the users were expecting the system to teach them what to do and correct them when they were wrong. For example, if a user performed a compression exam without click on the corresponding button or without a seed point, it seems they expected the system itself to tell them, instead of placing that cognitive burden on the user themselves. Therefore, if it is expected that a system such as this would be used mostly by novice users with little or no experience, more ‘user error catching’ routines would need to be implemented. 18 19  User commented on difficulty of using the trackball for the system, during the informal interview. The use of a touchscreen obviates the need to include ‘hotkeys’ on the button faces.  141  5.3.2  Clinical Evaluations  Two clinical experiments were performed for evaluating the DVT screening system. The first, pilot study, conducted between 2005 and 2006, provided information on patient populations, the approximate system sensitivity and specificity, was well as being useful for identifying technical issues. The second evaluation incorporated changes based on the results of the pilot study and usability evaluation. Pilot Study From the lengths of the scanned models in the pilot study, it is clear that the complete deep venous system is not being scanned. Even so, based on scanned vessel lengths and scan times, it is conceivable that screening the deep venous system (∼ 50 cm) may only take about 20 min using our system. If, in addition, scanning can be performed by a broad range of users, highly experienced staff could potentially see a decrease in workload – where they only scan patients positive for screening – with an increased detection rate. While some 3–D vessel models (34%) had slices with no compression data, half of these were located at the ends of the model, where we would expect some overlap between adjacent vessel scans. Only 6 (14.6%) of the models had empty slices in the centre of the model, and these would be considered incomplete scans. The scanning protocol would require that the latter segments be re-scanned in order to obtain vessel compressibility. The reasons for incomplete models in the pilot study were mainly because of failure of the contour detection algorithm, especially when vessels are completely compressed. In these cases, the feature that we want to detect disappears from the image and the contour detection algorithm either does not return a contour or detects an adjacent feature such as an artery, dark muscle or lymph node. Changes to the prototype system were made to address these problems. While the average TAR for healthy vessels lies below the optimal cutoff value of 55% (see Section 4.3.1)(t-test, p-value = 0, α = 0.05), the average TAR for diseased vessels does not. Indeed, only 1 of the 3 vessels has a TAR value indicating DVT. This results in a specificity of 92.1% and a sensitivity of 33.3% for the TAR measure in this data set. Slightly better performance was obtained when using the slope criteria. When using the specified cutoff value of -0.2 (see Section 4.3.1), half of the healthy vessels were incorrectly classified as probable for DVT, resulting in a specificity of 50%. On the other hand, all diseased vessels were correctly identified, and the mean value is greater than the cutoff of -0.2 (t-test, p-value = 0.105, α = 0.2), for a sensitivity of 100%. Improved specificity is obtained for the slope criteria if the mean slope value is used instead of the maximum value for each vessel model. That is, the resulting value for each criterion is obtained by averaging the slope values obtained for the scanned vessel segment. In this case, specificity improves to 89.5% but sensitivity drops to 66.7%. The main reasons for the inferior performance of the measures were outliers in the compression data and fitting a single line to two or more compression data sets. Again, changes were made to the prototype system to deal with these issues in the form of performing outlier detection on-the-fly. 142  Clinical Study The second study provided further information on system and vessel assessment performance, obtained while using the system in a clinical setting with potential users. The inability or difficulty with the visualization of patients sheds light on an important point. Most patients in the pilot study were outpatients and therefore able to move on their own, and only 3 of 23 had undergone recent surgery in the previous 2 weeks (2 knee replacement surgery, 1 ACL (anterior cruciate ligament) repair). In the final clinical exam, all patients had undergone surgery and there was a higher proportion of patients with medium and heavy build, generally indicating that legs were larger, and veins were deeper. Other factors such as swelling, immobility or other unidentified reasons contributed to the difficulty with visualization of the deep veins. This indicates that some of the basic assumptions for vessel detection and tracking, such as the fact that vessels will appear dark surrounded by brighter tissue20 , may need further generalization in order to detect vessels in all cases. The mean scan time increased three-fold from that obtained in the pilot study. This increase is due to two main factors. First, an additional assessment is being performed in the form of the flow evaluation, and extra time is needed in order to switch ultrasound modes and modify the settings. Second, the scans were performed by nurse practitioners who have no previous experience with ultrasound. While one of the goals of our system is to provide a user-independent approach, we realize that some time will be needed in order to become proficient in the use of the system. Indeed, the investigators observed a decrease in scan time as the study progressed. In addition, it is expected that further integration with the Sonix platform, in the form of detailed presets or other control, would reduce scan time. The sensitivity and specificity results from the individual compressibility criteria are very similar to the results obtained in the pilot study. Again, the slope criteria outperforms the Transverse Area Ratio, with a maximum sensitivity of 100%. Several interesting results were obtained with the calculation of the flow ratio. In particular, the maximum flow ratio values obtained (1.0 to 1.7) were much lower than those reported in the literature (1.8 to 4.8) [113]. Indeed, a flow ratio of 1.0 would indicate that there was no detected change in the flow when the patient flexed their ankle compared to when the patient lay at rest. The reasons for this are due to several factors. First, the type of flow information used was different, since the current vessel characterization system used pulsed wave (PW) Doppler generated flow images and the previous work used continuous wave (CW) Doppler to measure blood flow. Second, some settings such as the angle were not optimized during data collection. Finally, probe movement during the data collection also introduced variation into the measurements. However, good results were obtained using the flow ratio calculation, with a sensitivity and specificity of 100% and 66.7%, respectively. In addition, a significant difference was observed between flow ratios obtained from patients with negative DVT results and those with positive DVT results. It is expected that improvements to the system in the form of automated angle calculation 20 For teaching purposes (textbooks, videos) this is in fact the assumption. Large, swollen patients who have undergone recent surgery comprise a specific subpopulation to which these assumptions may not be applied as broadly.  143  (using location sensor measurements and known model location), detailed preset flow settings (or including control of the flow settings through the vessel characterization system interface), and the addition of flow tracking (flow is better seen on a longitudinal scan, where the current contour detection cannot be directly applied) would improve the specificity of this single measure. While the individual measures may not be as sensitive or specific as hoped, the combination of these measures yields very good results. When combining several of the most useful measures (based on the J value, Table 4.34) by calculating an optimal feature, a very significant increase is seen in the ability to separate scans that are positive from those that are negative for DVT. Indeed, while the individual features have J values ranging from 1.36 to 1.73 (only slightly better than chance), the optimal feature has a value of 80.4, with the positive and negative classes showing a very clear and pronounced separation. The implementation of this combined feature yields a system sensitivity of 100% and specificity of 75%.  5.3.3  Costs of a DVT Screening System  In order to assess the potential impact of a DVT screening system such as the one presented here, an estimate of the cost of the DVT screening system was elaborated. In addition, an estimate of the cost implications of using a DVT screening system was also generated. This is presented in Appendix C. Three comparisons of a proposed DVT screening system and its cost when compared to conventional screening are presented in Table C.4, Table C.5 and Table C.6. By taking into account possible savings based on the decrease in scan time, and the assumed difference in initial investment between the two methods, it can be seen in Table C.6 that substantial savings are possible if 10% to 20% of the patients scanned are sent for conventional examinations. It is reasonable to assume that this percentage of patients would be re-scanned. An interesting result in Table C.6 shows that for 50% specificity, the cost difference for 1,000 scanned patients is about $27,000. This is approximately equal to the estimated cost difference between a DVT screening system and a conventional ultrasound machine. Based on these numbers, scanning with either method will take the same amount of time, and the only difference in cost is the difference in the price of the equipment. Of course, the estimates presented in Table C.4, Table C.5 and Table C.6 depend on several factors, including the specificity of the system, the manner in which savings are calculated based on the decrease in scan time, and the amount of the initial investment needed for each of the different methods. Additionally, this analysis depends on the final price of a commercial version of a DVT screening system, especially the estimates of the number of patients to scan and the corresponding time in order to recover the initial investment by the health care provider. The estimates in Table C.5 - Table C.9 support the argument that with a reduced scan time, more patients can be screened for DVT and also help reduce the overall cost on the health care system by treating patients with DVT in earlier stages, instead of treating them once the disease has progressed and had more time to do damage. Section C.3.4 attempts to address the cost of less than perfect sensitivity for the DVT screening 144  system. From Table C.9 we can see that, simply by comparing the cost of screening patients for DVT plus the cost of detecting and treating the patients that were missed by the screening, it is less costly to permit the disease to progress. If we further analyze the repercussions of less than perfect sensitivity, depending on how we quantify the cost of patient complications and death, the gap between hypothetical systems with different sensitivity values increases. It should be noted that the cost of screening patients that are truly negative for DVT is not included in this Table. It is assumed that this would be a constant cost for all cases of sensitivity. Finally, an additional and important point regarding the DVT screening system should be made. For estimating the cost of scanning using a DVT screening system, it was assumed that an experienced ultrasound technician would perform the examination. In reality, the sensitivity of conventional screening can depend on the experience of the operator, and it may be difficult to obtain skilled technicians for scanning at all times. The proposed DVT screening system could overcome this dual problem of availability and variability with standardized measures and results, allowing for a greater pool of people to be used as operators. This, in turn, may also help reduce costs further.  5.4  Contributions  At the onset of this project, several ambitious goals were defined with the overall aim of creating a usable and accurate system that could detect and screen for DVT. While the overall aim would be the prefect detection of DVT in a range of patient populations, the system developed during this research addresses DVT detection in specific subpopulations and with specific users in mind. The main focus of this project has been to develop a screening system for DVT that is fast, accurate and user friendly, that could be used by health professionals with limited ultrasound experience, such as ward nurses, in order detect DVT. Such a system could subsequently reduce the complications that arise from this disease, such as PE, because of earlier diagnosis and treatment [134]. Recent estimates of deaths each year because of PE are 150,000 - 200,000 in the United States [125]. Current medical practice limits the number of patients that can be scanned for DVT, and a screening system such as that proposed here would be greatly beneficial in a setting such as an orthopaedic surgery ward, where it is known that 20% - 30% of patients who have undergone hip replacement surgery will develop DVT [27]. Relying on clinical signs or symptoms is known to be inaccurate, fewer than a third of patients actually present with classic DVT symptoms [48, 134], further justifying the need for a screening system. If one of the ultimate goals is the prevention of PE, then screening for DVT at the patient’s bedside would be extremely useful [55]. Approximately 80% of the emboli to the lungs arise from thrombi in the leg veins [57], yet less than 20% have signs and symptoms suggestive of DVT [28, 71]. A screening system that could identify even a fraction of patients with asymptomatic DVT before it progresses to PE would be extremely beneficial. The current approach has also followed the ‘single ultrasound’ examination strategy for the detection of DVT [27, 114, 141] that includes the strict standardization of the ultrasound exam145  ination protocol and the reduction of user variability, achieved by a detection system based on objective, numerical measures. Finally, the clinical evaluation of any developed system was also a goal of this research. While the overall aim of detecting all DVT is still incomplete and some suggestions for future work are outlined in the following Section, the current research has made substantial contributions to creating a DVT screening system. Specifically: • A functional system has been developed that makes it feasible to clinically test the hypothesis: “An ultrasound based system with numerical measures for DVT can be used to accurately screen patients for thrombus, where screening is performed at the bedside by nurses or others without extensive ultrasound experience or training.” In support of this: • A novel approach for using a geometric model in the measurement equation of an extended Kalman filter for contour detection and parameter estimation. This is the first time that an extended Kalman filter has been used with a model for vessel (or feature) detection in the measurement equation. Previous methods use either linear or extended Kalman filters with the dynamics described in the system equation [1, 3, 60]. In addition, this is the first time that multiple measurements – such as radial similarity and elastography data (Section 3.1.6) – have been used with the Kalman filter framework and a single model. In addition, other novel initialization and application dependent settings, such as angledependent covariance, the initial scan direction, and the use of a similarity measure to provide smoothness from a contour in one frame to the next, have also been presented. Various models were implemented (3–parameter ellipse, 5–parameter ellipse, ellipsoid) using the current approach and extensively tested and compared to the previous contour detection algorithms. Over the course of conducting this research, it has become clear that there are other applications in which such an algorithm could detect cavities or other hypoechoic regions, for example in the delineation of the prostate in ultrasound images [11, 12]. • The introduction of a new radial scanning approach with an increased step size that significantly speeds up parameter convergence when using a spatial extended Kalman filter for parameter estimation. All previous methods that use a star-Kalman approach [1, 2, 3, 60, 63, 64] (including those with extended Kalman filters) use a small angular step size. The experiments show that the same parameter estimation accuracy can be obtained with a three-fold increase in processing speed using this method. In real-time applications, any increase in speed is beneficial. • The use of a Kalman filter using a model with delayed measurements for tracking vessels in image sequences. While Kalman tracking has been extensively used in target tracking, this approach using delayed measurements has not been reported for image processing. Previous approaches [1, 60] used potentially unreliable measurements in the current time step, leading to larger errors. 146  • The first use of a sub-sampled approach for detecting locations of minimum brightness and of maximum similarity with a known mask was shown to be as accurate as calculating these measures using all data, but with a proportional decrease in computation time. The sum of absolutes is a known metric for gauging similarity in ultrasound [52, 54], but with the drawback of potentially high processing time. Again, the increase in speed that is gained using the sub-sampled approach is crucial for real-time tracking and detection applications. • A method for the identification of vessel bifurcations in 2–D ultrasound images using Bmode and colour flow information. While other methods exist for segmenting vascular trees and detecting vessel bifurcations, these rely on 3–D data sets [22] and/or do not necessarily construct a segmented model, rather they are visualizations of 3–D data [138]. No currently known approach addresses the specific problem encountered in this research regarding the detection of vessel bifurcations. In addition, this segmentation of the flow data can also be used for automated tracking initialization in the context of the developed vessel characterization system. • The implementation of a real-time vessel characterization system that incorporates procedures for a screening protocol for DVT. This system includes the integration of many complex systems into a single system, components such as the Ultrasonix Sonix R ultrasound machine, commercially available sensors, custom made pieces and a considerable amount of software with the described novel contour detection and tracking algorithms. This fully functional and stable system can be used to characterize vessels based on compressibility and blood flow, or for research and experimentation using elastography and vessel displacement information. • The combination of compressibility and flow measures from a vessel scan into a single feature that can differentiate between patients with and without DVT. All previous work has focused on a single type of measure (e.g. compressibility, flow, strain) and this is the first time that a formalized measure combining information from these different sources has been used. • The development of a usable interface for the vessel characterization system based on the known literature [40, 98, 140] and the evaluation of the usability of the system and interface with medical professionals, who comprise potential users of such a system. The learnability, memorability and user satisfaction with the system was evaluated, providing feedback and a baseline for any future user evaluations. • The clinical evaluation of the DVT screening system consisting of an initial pilot study and a final trial, both conducted at the University of British Columbia Hospital. The system was evaluated in a real world setting, and the evaluation provided information on the performance of the system compared to conventional procedures. Previous work [60] was limited to the examination of healthy individuals in a laboratory setting using a limited system. Several peer–reviewed publications have arisen from this work, including two journal papers: • Guerrero et al , System for Deep Venous Thrombosis Detection using Objective Compression Measures, IEEE Transactions on Biomedical Engineering [66], which presents an overview of 147  the screening system and it’s components, including the compressibility measures, • Guerrero et al , Real-time Vessel Segmentation and Tracking for Ultrasound Imaging Applications, IEEE Transactions on Medical Imaging [67], which presents the 3–parameter ellipse model contour detection and vessel tracking algorithms in detail, and three conference presentations: • Guerrero, Real-time vessel tracking for deep venous thrombosis screening system, Intelligent Systems (IS) 2004 [61], which describes the vessel compresibility assessment, • Guerrero et al , Fast Screening System for Deep Vein Thrombosis, 29th Conference of the Canadian Medical and Biological Engineering Society [65], which presents the results from the clinical pilot study, and • Guerrero, Deep vein thrombosis screening and vessel compressibility characterization, Intelligent Systems (IS) 2008 [62], which presents details on the 5–parameter ellipse model for contour detection, staggered radial scanning, and vessel bifurcation detection.  5.5  Future Work  As with many ambitious projects, areas that require further work still remain. This Section includes some recommendations for future work. In general, the goals of any future research that follows the work presented here would be to consolidate the results obtained with the vessel characterization system and to address the weak points of the system that have been identified. In other words, a substantial and statistically significant dataset is needed to fully determine the sensitivity and specificity of this system, using the proposed measures, a combination of these, or new measures based on the available data. Limitations of conventional B-mode ultrasound were made evident in the final clinical study from the inability to visualize some patients, and reported difficulties by sonographers as well in locating the veins in the deep venous system. This is reflected in our system as a current weakness but other imaging modalities, such as the emerging strain imaging or elastography, may provide additional information that could lead to a vessel characterization system that overcomes even these conventional limitations. The use of strain images and corresponding correlation maps may allow the development of new and complementary vessel assessment measures that could improve system performance in the detection of DVT. Investigators should therefore focus on these areas in future efforts. Notwithstanding the known limitations with ultrasound, improvements to the current approach using compression measures are possible, including refinement to the contour detection and vessel tracking, processing all the compression data (during compression and during release, by for example buffering the data obtained during release and processing it in reverse order), and combining different compressibility measures to improve the system’s ability to distinguish between patients positive and negative for DVT. 148  Time and effort should also be devoted to streamlining the scanning approach from the user’s point of view. Questions such as • How can the system be more intuitive to use? • How can user variability be reduced? • How can users accurately scan patients faster? should also be addressed, including experimental studies to validate results. A further extension of the work presented here would be to examine the potential of the vessel characterization system as a replacement for the current DVT diagnostic standard, compression ultrasound. While the current approach has been to obtain a system to perform a fast screening examination, the use of the standardized compressibility and DVT measures could be incorporated into a system that requires a more detailed examination and improve upon current and established detection methods for DVT. Other specific recommendations are: • Throughout the experiments, the measures used to detect DVT have been evaluated independently. A promising approach for DVT detection is to combine some or all of these measures into a single DVT measure, with higher sensitivity and specificity. Indeed, the evaluations presented at the end of Chapter 4 indicate that perfect detection may be possible. The extension of this approach to include measures from different sources (such as strain data) and an in-depth validation is needed. • Throughout the development of the contour detection and tracking, several alternate applications either became apparent or were suggested. Tracking local deformations of the vessel lumen from frame to frame may provide additional information about the local stiffness of the vessel. This information, coupled with other data such as elastographs and blood flow within the vessel, may be useful for detecting arterial hardening. The feasibility of using these tools to detect atherosclerosis should be explored. Additionally, similar tissue characterization applications using the presented sensorized probe are easy to envision. An interesting extension of this system is as a platform for the real time evaluation of elastic properties of different anatomical structures based on feature extraction and the sensorized probe. • It is believed that tracking models that take into account the location of multiple features in a single image frame (such as the location of the vein and of the artery) are a promising approach for increased reliability of the vessel tracking. While this type of model has been presented and is used in the system, the inclusion of detailed information to define a tracking model – e.g. an artery should be above a vein in the image, both vessels cannot occupy the same space, during compression an artery and a vein (and their centre points) will behave differently – may further increase the reliability of the vessel tracking algorithms and should be developed further. 149  • The main limitation for adequate detection in real-time of vessel bifurcations is the prohibitive drop in frame rate of the ultrasound data when flow data is being acquired. It is conceivable that detecting vessel bifurcations using color flow imaging could become practical if the timing of the ultrasound acquisition and the selected region of interest could be quickly updated. The current Ultrasonix software development kits may allow for this possibility, but extensive development may be required. • Because of practical and time limitations, the system performance for calf has not been fully evaluated. Clinical data to determine the ability to detect calf DVT is needed.  150  Bibliography [1] P. Abolmaesumi, S. E. Salcudean, W. H. Zhu, M. R. Sirouspour, and S. P. DiMaio. ImageGuided Control of a Robot for Medical Ultrasound. IEEE Trans. Robot. Automat., 18(1), 2002. [2] P. Abolmaesumi and M. R. Sirouspour. An Interacting Multiple Model Probabilistic Data Association Filter for Cavity Boundary Extraction from Ultrasound Images. IEEE Trans. Med. Imag., 23(6):772–784, 2004. [3] P. Abolmaesumi, M. R. Sirouspour, and S. E. Salcudean. Real-Time Extraction of Carotid Artery Contours from Ultrasound Images. In CBMS, pages 181 – 186, Houston, USA, June 2000. IEEE Computer Society. [4] S. Aglyamov, A. R. Skovoroda, J. M. Rubin, M. O’Donnell, and S. Y. Emelianov. Modelbased reconstructive elasticity imaging of deep venous thrombosis. IEEE Trans. Ultrason., Ferroelect., Freq. Contr., 51(5):521–531, May 2004. [5] The Diagnostic Approach to Acute Venous Thromboembolism. Clinical Practice Guideline. Am J Respir Crit Care Med, 160(3):1043–1066, 1999. [6] E. J. Andrews, Jr. and A. C. Fleischer. Sonography for Deep Venous Thrombosis: Current and Future Applications. Ultrasound Q, 21(4):213–225, Dec. 2005. [7] A. M. Asbeutah, A. Z.Riha, J. D. Cameron, and B. P. McGrath. Five-year outcome study of deep vein thrombosis in the lower limbs. J Vasc Surg, 40(6):1184–1189, Dec. 2004. [8] M. Atri, M. J. Herba, C. Reinhold, C. Leclerc, S. Ye, F. F. Illescas, and P. M. Bret. Accuracy Of Sonography in the Evaluation of Calf Deep Vein Thrombosis in both Postoperative Surveillance and Symptomatic Patients. AJR Am J Roentgenol, 166:1361–1367, 1996. [9] Australia’s health ah96-x04.html.  1996.  http://www.aihw.gov.au/publications/health/ah96/  [10] S. Badiei. Prostate segmentation in ultrasound images using image warping and ellipsoid fitting. Master’s thesis, University of British Columbia, 2007. [11] S. Badiei, S. E. Salcudean, J. Varah, and W. J. Morris. Prostate segmentation in 2d ultrasound images using image warping and ellipse fitting. In MICCAI, 2006. [12] S. Badiei, S. E. Salcudean, J. Varah, N. Usmani, N. Chng, A. Kruck, I. Spadinger, and W. J. Morris. 3d prostate segmentation in trans-rectal ultrasound images using image warping and ellipsoid fitting. In American Brachytherapy Society Conference, 2007. [13] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan. Estimation with Applications to Tracking and Navigation. John Wiley & Sons, Inc., New York, USA, 2001. [14] D. Becker, E. G¨ unter, R. Strauss, K. Cidlinsky, B. Tomandl, D. Kalden-Nemeth, D. Neureiter, A. Lang, and E. G. Hahn. Color Doppler Imaging Versus Phlebography in the Diagnosis of Deep Leg and Pelvic Vein Thrombosis. Journal of Ultrasound in Medicine, 16:31–37, 1997. 151  [15] D. J. Berry. Surveillance for Venous Thromboembolic Disease After Total Knee Arthroplasty. Clin Orthop Relat Res, 392:257–266, Nov. 2001. [16] A. Blake and A. Yuille, editors. Active Vision, chapter Tracking with Kalman Snakes, D. Terzopoulos and R. Szeliski. The MIT Press, 1992. [17] E. I. Bluth. Leg Swelling with Pain or Edema. In A Special Course in Ultrasound, pages 99–105. RSNA, 1996. [18] D. Boukerroui, O. Basset, A. Baskurt, and G. Gimenez. A multiparametric and multiresolution segmentation algorithm of 3-D ultrasonic data. IEEE Trans. Ultrason., Ferroelect., Freq. Contr., 48(1):64–77, Jan. 2001. [19] J. Brown. Assessment of pretest risk for venous thromboembolic disease. Emerg Med Clin North Am, 19(4):861–868, Nov 2001. [20] Candian dollar exchange rate. http://www.bankofcanada.ca/en/rates/exchange-avg. html. [21] V. Chalana and Y. Kim. A Methodology for Evaluation of Boundary Detection Algorithms on Medical Images. IEEE Trans. Med. Imag., 16(5):642–652, 1997. [22] C. Chen, T. Poepping, J. Beech-Brandt, S. Hammer, R. Baldock, B. Hill, P. Allan, W. Easson, and P. Hoskins. Segmentation of arterial geometry from ultrasound images using balloon models. In IEEE International Symposium on Biomedical Imaging: Macro to Nano, volume 2, pages 1319 – 1322. IEEE, April 2004. [23] D. Chen, R. Chang, W. Wu, W. K. Moon, and W. Wu. 3D Breast Ultrasound Segmentation Using Active Contour Model. Ultrasound Med Biol, 29(7):1017–1026, 2003. [24] C. K. Chui and G. Chen. Kalman Filtering with Real-Time Applications. Springer, 3rd edition, 1999. [25] R. M. Comeau, A. F. Sadikot, A. Fenster, and T. M. Peters. Intraoperative Ultrasound for Guidance and Tissue Shift Correction in Image-Guided Neurosurgery. Medical Physics, 27(4):787–800, 2000. [26] J. Constants, C. Boutinet, L. R. Salmi, J. C. Saby, M. L. Nelzy, P. Baudouin, F. Sampoux, J. M. Marchand, C. Boutami, V. Dehant, S. Pulci, J. P. Gauthier, V. Cacareigt-Bourdenx, D. Barcat, and C. Conri. Comparison of Four Clinical Prediction Scores for the Diagnosis of Lower Limb Deep Venous Thrombosis in Outpatients. The American Journal of Medicine, 115:436–440, 2003. [27] J. Cornuz, S. D. Pearson, and J. F. Polak. Deep Venous Thrombosis: Complete Lower Extremity Venous US Evaluation in Patients without Known Risk Factors - Outcome Study. Radiology, 211:637–641, 1999. [28] J. J. Cranley, A. J. Canos, and W. J. Sull. The Diagnosis of Deep Venous Thrombosis: Fallibility of Clinical Symptoms and Signs. Archives of Surgery, 111:34–36, 1976. [29] J. J. Cronan. Venous Thromboembolic Disease: The Role of US. Radiology, 186:619–630, 1993. [30] J. J. Cronan, G. Dorfman, F. Scola, B. Schepps, and J. Alexander. Deep Venous Thrombosis: US Assessment Using Vein Compression. Radiology, 162:191–194, 1987. 152  [31] J. J. Cronan, G. S. Dorfman, and J. Grusmark. Lower-Extremity Deep Venous Thrombosis: Further Experience With And Refinements Of US Assessment. Radiology, 168:101–107, 1988. [32] J. J. Cronan and V. Leen. Recurrent Deep Venous Thrombosis: Limitations of US. Radiology, 170:739–742, 1989. [33] M. A. Crowther, D. J. Cook, L. E. Griffith, P. J. Devereaux, C. C. Rabbat, F. J. Clarke, N. Hoad, E. McDonald, M. O. Meade, G. H. Guyatt, et al. Deep venous thrombosis: clinically silent in the intensive care unit. J Crit Care, 20(4):334–340, Dec. 2005. [34] M. Dauzat, J.-P. Laroche, G. Deklunder, J. Ayoub, I. Qu´er´e, and C. J. Fran¸cois-Michel Lopez. Diagnosis of acute lower limb deep venous thrombosis with ultrasound: Trends and controversies. J Clin Ultrasound, 25(7):343–358, Sept. 1997. [35] J. C. B. David S. Bell and R. J. Eckersley. Segmentation and Analysis of Colour Doppler Images of Tumour Vasculature. Ultrasound Med Biol, 21(5):635–647, 1995. [36] C. L. de Korte, E. I. C´espedes, A. F. W. van der Steen, and C. T. Lanc´ee. Intravascular Elasticity Imaging using Ultrasound: Feasibility Studies in Phantoms. Ultrasound Med Biol, 23(5):735–746, 1997. [37] P. T. Diamond and S. N. Macciocchi. Predictive Power of Clinical Symptoms in Patients with Presumptive Deep Venous Thrombosis. American Journal of Physical Medicine & Rehabilitation, 76(1):49–51, 1997. [38] D. B. Downey and A. Fenster. Vascular Imaging with a Three-Dimensional Power Doppler System. American Journal of Roentgenology, 165:665–668, 1995. [39] M. M. Doyley, C. L. de Korte, F. Mastik, S. G. Carlier, A. F. W. van der Steen, P. W. Serruys, and N. Bom. An Automated Approach for Clinical Intravascular Ultrasonic Palpation. In Ultrasonics Symposium, pages 1779 – 1782. IEEE, 2000. [40] J. S. Dumas and J. C. Redish. A Practical Guide to Usability Testing. Intellect Books, revised edition, 1999. [41] A. Elias, A. Cadene, M. Elias, J. Puget, J. L. Tricoire, C. Colin, D. Lefebvre, H. Rousseau, and F. Joffre. Extended Lower Limb Venous Ultrasound for the Diagnosis of Proximal and Distal Vein Thrombosis in Asymptomatic Patients after Total Hip Replacement. Eur J Vasc Endovasc Surg, 27(4):438–444, Apr. 2004. [42] S. Y. Emelianov, X. Chen, M. O’Donnell, B. Knipp, D. Myers, T. W. Wakefield, and J. M. Rubin. Triplex Ultrasound: Elasticity Imaging to Age Deep Venous Thrombosis. Ultrasound Med Biol, 28(6):757–767, 2002. [43] FDA Device Advice. http://www.fda.gov/cdrh/devadvice/313.html. [44] FDA search. CFRSearch.cfm.  http://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfcfr/  [45] R. Feld. Pulmonary Embolism Caused by Venous Compression Ultrasound Examination. J Vasc Surg, 17(2):450, 1993. in Letters to the Editor. [46] A. Fenster and D. B. Downey. 3-D Ultrasound Imaging: A Review. IEEE Eng. Med. Biol. Mag., 5:41–51, Nov./Dec. 1997. 153  [47] A. Fenster, D. Lee, S. Sherebrin, R. Rankin, and D. B. Downey. Three-Dimensional Ultrasound Imaging of the Vasculature. Ultrasonics, 36:629–633, 1998. [48] G. Fern´andez-Canton, I. Lopez Viadaur, F. Mu˜ noz, M. A. Anto˜ nana, F. Uresandi, and J. Calonge. Diagnostic Utility of Color Doppler Ultrasound in Lower Limb Deep Vein Thrombosis in Patients with Clinical Suspicion of Pulmonary Thromboembolism. European Journal of Radiology, 19:50–55, 1994. [49] J. D. Fraser and A. D. R. Deep Venous Thrombosis: Recent Advances and Optimal Investigation with US. Radiology, 211:9–24, 1999. [50] M. G. Frederick, B. S. Hertzberg, M. A. Kliewer, E. K. Paulson, J. D. Bowie, K. J. Lalouche, D. M. DeLong, and B. A. Carroll. Can the US examination for lower extremity deep venous thrombosis be abbreviated? A prospective study of 755 examinations. Radiology, 199(1):45– 47, 1996. [51] N. Friedland and D. Adam. Automatic Ventricular Cavity Boundary Detection from Sequential Ultrasound Images Using Simulated Annealing. IEEE Trans. Med. Imag., 8(4):344–353, 1989. [52] B. H. Friemel, L. N. Bohs, and G. E. Trahey. Relative Performance of Two-Dimensional Speckle-Tracking Techniques: Normalized Correlation, Non-Normalized Correlation and Sum-Absolute-Difference. In Proc. (IEEE) Ultrasonics Symposium, pages 1481 – 1484, 1995. [53] D. Gaitini. Current approaches and controversial issues in the diagnosis of deep vein thrombosis via duplex Doppler ultrasound. J Clin Ultrasound, 34(6):289–297, Jul./Aug. 2006. [54] C. M. Gallippi, L. N. Bohs, M. Anderson, A. Congdon, and G. E. Trahey. Lateral Blood Velocity Measurement in the Carotid Artery Via Speckle Tracking. In Proc. (IEEE) Ultrasonics Symposium, pages 1451 – 1455, 2001. [55] A. J. P. Goddard, S. Chakravert, and J. Wright. Computer Assisted Strain-gauge Plethysmography is a Practical Method of Excluding Deep Venous Thrombosis. Clin Radiol, 56(1):30–34, Jan. 2001. [56] S. Goodacre, M. Stevenson, A. Wailoo, F. Sampson, A. Sutton, and S. Thomas. How should we diagnose suspected deep-vein thrombosis? QJM, 99(6):377–388, 2006. [57] R. H. Gottlieb, S. L. Voci, L. Syed, C. Shyu, P. J. Fultz, D. J. Rubens, J. G. Strang, N. Carson, W. J. DiGrazio, and C. W. Francis. Randomized Prospective Study Compariing Routine Versus Selective Use of Sonography of the Complete Calf in Patients with Suspected Deep Venous Thrombosis. American Journal of Roentgenology, 180:241–245, 2003. [58] R. H. Gottlieb, J. Widjaja, S. Mehra, and W. B. Bobinette. Clinically Important Pulmonary Emboli: Does Calf Vein US Alter Outcomes? Radiology, 211:25–29, 1999. [59] M. T. Grebe. Compression Sonography for the Exclusion of Deep Vein Thrombosis: Can I trust my eyes? Thromb Haemost, 89(2):206–207, 2003. Editorial. [60] J. Guerrero. Improved interface for deep venous thrombosis screening using B-Mode ultrasound. Master’s thesis, University of British Columbia, Vancouver, 2002. [61] J. Guerrero. Real-time vessel tracking for deep venous thrombosis screening system. In Intelligent Systems (IS), 2004. 154  [62] J. Guerrero. Deep vein thrombosis screening and vessel compressibility characterization. In Intelligent Systems (IS), 2008. [63] J. Guerrero, S. E. Salcudean, J. A. McEwen, B. A. Masri, and S. Nicolaou. Deep Venous Thrombosis Screening System Using Numerical Measures. In 25th Annual International Conference of the IEEE EMBS, pages 894–897. IEEE/EMBS, September 2003. [64] J. Guerrero, S. E. Salcudean, J. A. McEwen, B. A. Masri, and S. Nicolaou. MeasurementBased Deep Venous Thrombosis Screening System. In 6th Annual International Conference on Medical Image Computing & Computer Assisted Intervention, pages 214–221. MICCAI, November 2003. [65] J. Guerrero, S. E. Salcudean, J. A. McEwen, B. A. Masri, and S. Nicolaou. Fast Screening System for Deep Vein Thrombosis. In 29th Conference of the Canadian Medical and Biological Engineering Society, Vancouver, BC, June 2006. CMBES. [66] J. Guerrero, S. E. Salcudean, J. A. McEwen, B. A. Masri, and S. Nicolaou. System for Deep Venous Thrombosis Detection using Objective Compression Measures. IEEE Trans. Biomed. Eng., 53(5):845–854, 2006. [67] J. Guerrero, S. E. Salcudean, J. A. McEwen, B. A. Masri, and S. Nicolaou. Real-time Vessel Segmentation and Tracking for Ultrasound Imaging Applications. IEEE Trans. Med. Imag., 26(8):1079–1090, 2007. [68] J. R. Harding. Investigating Deep Vein Thrombosis With Infrared Imaging. IEEE Eng. Med. Biol. Mag., July/August 1998. [69] P. Hassenpflug, R. W. Prager, G. M. Treece, and A. H. Gee. Speckle classification for sensorless freehand 3-D ultrasound. Ultrasound Med Biol, 31(11):1499–1508, Nov. 1998. [70] M. Hern´andez-Hoyos, A. Anwander, M. Orkisz, J. P. Roux, P. Douek, and I. E. Magnin. A Deformable Vessel Model with Single Point Initilization for Segmentation, Quantification and Visualization of Blood Vessels in 3D MRA. In MICCAI 2000, pages 735 – 745, 2000. [71] J. Hirsh and J. Hoak. Management of Deep Vein Thrombosis and Pulmonary Embolism. Circulation, 93:2212–2245, 1996. [72] J. Hirsh and A. Y. Y. Lee. How we diagnose and treat deep vein thrombosis. Blood, 99(9):3102–3110, May 2002. [73] S. W. Hughes, T. J. D’Arcy, D. J. Maxwell, W. Chiu, A. Milner, J. E. Saunders, and R. J. Sheppard. Volume Estimation From Multiplanar 2D Ultrasound Images Using A Remote Electromagnetic Position And Orientation Sensor. Ultrasound Med Biol, 22(5):561–572, 1996. [74] The Visualization Toolkit, 2005. http://www.vtk.org/. [75] O. A. Ilahi, J. Reddy, and I. Ahmad. Deep Venous Thrombosis After Knee Arthroscopy: A Meta-analysis. Arthroscopy, 21(6):727–730, June 2005. [76] Intel Compiler, 2006. http://www.intel.com/cd/software/products/asmo-na/eng/ compilers/. Visit http://www.intel.com/ and follow links to Software and Compilers if link is broken. [77] Intel Math Kernel Library (MKL), 2006. http://www.intel.com/cd/software/products/ asmo-na/eng/perflib/mkl/. Visit http://www.intel.com/ and follow links to Software and Performance Libraries if link is broken. 155  [78] R. Iorio, S. Dhupar, W. L. Healy, and E. Wong. Routine Duplex Ultrasound Screening after TKA is Not Necessary. Clin Orthop Relat Res, 452:171–174, 2006. [79] M. C. H. Janssen, H. Wollersheim, I. R. O. Novakova, F. M. J. Heystraten, W. N. J. C. van Asten, and T. Thien. Diagnosis Of Deep Vein Thrombosis, An Overview. Neth J Med, 48:109–121, 1996. [80] J. A. Jensen. Field: A Program for Simulating Ultrasound Systems. Med Biol Eng Comput, 34:351–353, 1996. Supplement 1, Part 1. [81] J. M. Johanning, D. P. Franklin, D. D. Thomas, and J. R. Elmore. D-dimer and calf circumference in the evaluation of outpatient deep venous thrombosis. J Vasc Surg, 36(5):877–880, Nov. 2002. [82] V. V. Kakkar, C. T. Howe, C. Flanc, and M. B. Clarke. Natural History Of Postoperative Deep-Vein Thrombosis. Lancet, 2:230–232, 1969. [83] M. Kass, A. Witkin, and D. Terzopoulos. Snakes: Active Contour Models. International Journal of Computer Vision, pages 321–331, 1988. [84] J. Kelly, A. Rudd, R. Lewis, and B. Hunt. Screening for subclinical deep-vein thrombosis. QJM, 94(10):511–519, 2001. [85] I. Lausen, R. Jensen, P. Wille-Jørgensen, L. N. Jørgensen, M. S. Rasmussen, K. M. Lyng, M. Anderson, and H. O. Raaschou. Colour Doppler Flow Imaging Ultrasonography versus Venography as Screening Method for Asymptomatic Postoperative Deep Venous Thrombosis. European Journal of Radiology, 20:200–204, 1995. [86] M. J. Ledesma-Carbayo, J. Kybic, M. Desco, A. Santos, M. S’uhling, P. Hunziker, and M. Unser. Spatio-temporal non-rigid registration for ultrasound cardiac motion estimation. IEEE Trans. Med. Imag., 24(9):1113–1126, Sept. 2005. [87] A. W. Lensing, C. I. Doris, F. P. McGrath, A. Cogo, M. J. Sabine, J. Ginsberg, P. Prandoni, A. G. Turpie, and J. Hirsh. A comparison of compression ultrasound with color Doppler ultrasound for the diagnosis of symptomless postoperative deep vein thrombosis. Arch Intern Med, 157(7):765–768, Apr. 1997. [88] D. F. Leotta, P. R. Detmer, and R. W. Martin. Performance of a Miniature Magnetic Position Sensor for Three-Dimensional Ultrasound Imaging. Ultrasound Med Biol, 23(4):597–609, 1997. [89] T. Locker, S. Goodacre, F. Sampson, A. Webster, and A. J. Sutton. Meta-analysis of plethysmography and rheography in the diagnosis of deep vein thrombosis. Emerg Med J, 23(8):630– 635, Aug. 2006. [90] M. E. Lockhart, H. I. Sheldon, and M. L. Robbin. Augmentation in Lower Extremity Sonography for the Detection of Deep Venous Thrombosis. AJR Am J Roentgenol, 184(2):419–422, 2005. [91] J. J. Mai, J. K. Tsuo, C. Pellot-Barakat, W. J. Hornof, C. Kargel, and M. F. Insana. Vascular Compliance Using Elasticity Imaging. In Ultrasonics Symposium, pages 1577–1580. IEEE, 2001. [92] F. Mao, J. Gill, D. Downey, and A. Fenster. Segmentation of Carotid Artery in Ultrasound Images: Method Development and Evaluation Technique. Medical Physics, 27(8):1961–1970, 2000. 156  [93] J. J. Michiels, A. Gadisseur, M. van der Planken, W. Schroyens, Z. Berneman, M. De Maeseneer, J. Hermsen, and P. Trienekens. Diagnosis of deep vein thrombosis: how many tests do we need? Acta Chir Belg, 105(1):16–25, 2005. [94] J. J. Michiels, A. Gadisseur, M. van der Planken, W. Schroyens, M. De Maeseneer, J. T. Hermsen, P. H. Trienekens, H. Hoogsteden, and P. M. P. Pattynama. Different accuracies of rapid enzyme-linked immunosorbent, turbidimetric, and agglutination d-dimer assays for thrombosis exclusion: Impact on diagnostic work-ups of outpatients with suspected deep vein thrombosis and pulmonary embolism. Semin Thromb Hemost, 32(7):678–693, 2006. [95] M. Mitsuishi, S. Warisawa, T. Tsuda, T. Higuchi, N. Koizumi, H. Hashizume, and K. Fujiwara. Remote Ultrasound Diagnostic System. In International Conference On Robotics And Automation, volume 2, pages 1567–1574. IEEE, 2001. [96] A. Mumme, W. Heinen, B. Geier, W. Maatz, L. Barbera, and G. Walterbusch. Regional hyperthermic fibrinolytic perfusion after unsuccessful venous thrombectomy of extensive deep venous thrombosis. J Vasc Surg, 36(6):1219–1224, Dec. 2002. [97] J. Nerney Welch, J. A. Johnson, M. R. Bax, R. Badr, and R. Shahidi. A Real-time Freehand 3D Ultrasound System for Image-guided Surgery. In Ultrasonics Symposium, pages 1601– 1604. IEEE, 2000. [98] J. Nielsen. Usability Enginneering. Academic Press, Inc., 1993. [99] J. A. Noble and D. Boukerroui. Ultrasound Image Segmentation: A Survey. IEEE Trans. Med. Imag., 25(8):987–1010, August 2006. [100] C. Pellot, I. Bloch, A. Herment, and F. Sureda. An Attempt to 3D Reconsruct Vessel Morphology From X-Ray Projections and Intravascular Ultrasounds Modelling and Fusion. Computerized Medical Imaging and Graphics, 20(3):141–151, 1996. [101] S. J. Perlin. Pulmonary Embolism during Compression US of the Lower Extremity. Radiology, 184:165–166, 1996. [102] N. Perone et al. Comparison Of Four Strategies For Diagnosing Deep Vein Thrombosis: A Cost-Effectiveness Analysis. The American Journal of Medicine, 110(1):33–40, 2001. [103] A. Petrie and C. Sabin. Medical Statistics at a Glance. Blackwell Science Ltd., 2000. [104] P. A. Picot, D. W. Rickey, R. Mitchell, R. N. Rankin, and A. Fenster. Three-Dimensional Colour Doppler Imaging. Ultrasound Med Biol, 19(2):95–104, 1993. [105] P. Prandoni and P. Mannuccio. Deep-Vein Thrombosis of the Lower Limbs: Diagnosis and Management. Bailliere’s Clinical Haematology, 12(3):533–554, 1999. [106] R. Puls, N. Hosten, J. S. B¨ock, J. H. Oellinger, A. J. Lemke, M. Gutberlet, K. Holz, and R. Felix. Signal-Enhanced Color Doppler Sonography of Deep Venous Thrombosis in the Lower Limbs and Pelvis. Journal of Ultrasound in Medicine, 18:185–190, 1999. [107] D. F. Riddle. Analytic Geometry. Wadsworth Publishing Company, 3 edition, 1982. [108] B. J. Robinson, P. J. L. Kesteven, and S. T. Elliott. The role of strain gauge plethysmography in the assessment of patients with suspected deep vein thrombosis. Br J Haematol, 118(2):600–603, Aug. 2002. 157  [109] S. C. Rose, W. J. Zwiebel, and F. Miller. Distribution of Acute Lower Extremity Deep Ve