Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Simultaneous analysis of 2D echo views for left atrial segmentation and disease quantification Allan, Gregory 2015

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

Item Metadata


24-ubc_2016_february_allan_gregory.pdf [ 20.85MB ]
JSON: 24-1.0222997.json
JSON-LD: 24-1.0222997-ld.json
RDF/XML (Pretty): 24-1.0222997-rdf.xml
RDF/JSON: 24-1.0222997-rdf.json
Turtle: 24-1.0222997-turtle.txt
N-Triples: 24-1.0222997-rdf-ntriples.txt
Original Record: 24-1.0222997-source.json
Full Text

Full Text

Simultaneous Analysis of 2D Echo Views for Left AtrialSegmentation and Disease QuantificationbyGregory AllanB.Sc. Biomedical Computing Honours, Queen’s University, 2013A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFMaster of Applied ScienceinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Electrical and Computer Engineering)The University of British Columbia(Vancouver)December 2015c Gregory Allan, 2015AbstractWe propose a joint information framework for automatic analysis of 2D echocar-diography (echo) data. The analysis combines a priori images, their segmentationsand patient diagnostic information within a unified framework to determine variousclinical parameters, such as cardiac chamber volumes, and cardiac disease labels.The main idea behind the framework is to employ joint Independent ComponentAnalysis of both echo image intensity information and corresponding segmentationlabels to generate models that jointly describe the image and label space of echopatients on multiple apical views jointly, instead of independently. These modelsare then both used for segmentation and volume estimation of cardiac chamberssuch as the left atrium and for detecting pathological abnormalities such as mitralregurgitation. We validate the approach on a large cohort of echos obtained from6,993 studies. We report performance of the proposed framework in estimationof the left-atrium volume and diagnosis of mitral-regurgitation severity. A corre-lation coefficient of 0.87 was achieved for volume estimation of the left atriumwhen compared to the clinical report. Moreover, we classified patients that sufferfrom moderate or severe mitral regurgitation diagnosis with an average accuracy of82%. Using only B-Mode echo information to automatically derive these clinicalparameters, there is potential for this approach to be used clinically.iiPrefaceThis thesis resulted from the collaboration between multiple researchers and isprimarily based on a pending journal submission. The contribution of the authorwas in developing, implementing, evaluating the presented framework and cre-ated a local database of relevant data. Ethical approval for conducting the projecttitled, ”Information Intelligence for Precision Cardiac Ultrasound Imaging”, hasbeen provided by the Vancouver Coastal Health Research Ethics Board, certificatenumbers: H13-02370.Saman Nouranian, Alexander Seitel, Maryam Mirian, Robert Rohling and Pu-rang Abolmaesumi helped with their valuable suggestions in improving the method-ology. Teresa Tsang, John Jue and Ken Gin collected the echocardiography datasets.Dale Hawley and Jody Swift assisted in obtaining ethics and obtaining access tothe echocardiography data.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiGlossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Proposed Framework . . . . . . . . . . . . . . . . . . . . . . . . 51.1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . 51.2 Structure of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 72 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.1 Anatomy of the Heart . . . . . . . . . . . . . . . . . . . . . . . . 112.2 Ultrasound Images . . . . . . . . . . . . . . . . . . . . . . . . . 122.2.1 B-Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.2.2 Colour Doppler . . . . . . . . . . . . . . . . . . . . . . . 132.3 Echocardiography . . . . . . . . . . . . . . . . . . . . . . . . . . 132.3.1 Sonographer Workflow . . . . . . . . . . . . . . . . . . . 142.3.2 Computer Workflow . . . . . . . . . . . . . . . . . . . . 172.3.3 Cardiologist Workflow . . . . . . . . . . . . . . . . . . . 19iv2.4 Echocardiography Diagnosis . . . . . . . . . . . . . . . . . . . . 192.4.1 Left Atrium Dynamics . . . . . . . . . . . . . . . . . . . 192.4.2 Mitral Regurgitation . . . . . . . . . . . . . . . . . . . . 222.4.3 Diastolic Function . . . . . . . . . . . . . . . . . . . . . 252.5 Related Segmentation Works . . . . . . . . . . . . . . . . . . . . 272.5.1 Boundary-Driven & Region-Based Segmentation Techniques 282.5.2 Model Fitting Techniques . . . . . . . . . . . . . . . . . 302.5.3 Radio Frequency Techniques . . . . . . . . . . . . . . . . 342.6 Commercial Products . . . . . . . . . . . . . . . . . . . . . . . . 353 Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.1 Ethics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.2 Echocardiography Data . . . . . . . . . . . . . . . . . . . . . . . 393.2.1 Echo Lab Software . . . . . . . . . . . . . . . . . . . . . 393.2.2 Data Acquisition and Pre-Processing . . . . . . . . . . . . 413.3 Patient Measurements . . . . . . . . . . . . . . . . . . . . . . . . 423.3.1 FileMaker Software . . . . . . . . . . . . . . . . . . . . 433.3.2 Data Acquisition & Data Pre-processing . . . . . . . . . . 444 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.1 View Preprocesing . . . . . . . . . . . . . . . . . . . . . . . . . 484.1.1 ROI Localization . . . . . . . . . . . . . . . . . . . . . . 494.2 LA Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . 504.2.1 Training: Patient Grouping . . . . . . . . . . . . . . . . . 514.2.2 Training: Joint Model Generation . . . . . . . . . . . . . 514.2.3 Training: Segmentation SVM Classification . . . . . . . . 534.2.4 Testing: Left Atrium (LA) Segmentation Based on EachJoint Independent Component Analysis (jICA) Model . . . 534.2.5 Testing: Final LA Segmentation . . . . . . . . . . . . . . 544.3 Volume Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 544.4 Disease SVM Classifier . . . . . . . . . . . . . . . . . . . . . . . 544.5 Diastolic Function . . . . . . . . . . . . . . . . . . . . . . . . . . 55v5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585.1 Framework Parameters . . . . . . . . . . . . . . . . . . . . . . . 595.2 Apical View Classification . . . . . . . . . . . . . . . . . . . . . 595.3 ROI Localization . . . . . . . . . . . . . . . . . . . . . . . . . . 605.4 Segmentation and Volume Estimation of the LA . . . . . . . . . . 605.5 MR Diagnostic Labels . . . . . . . . . . . . . . . . . . . . . . . 605.6 Diastolic Dysfunction and Left Ventricle (LV) Filling Pressure . . 605.7 Time and Computational Complexity . . . . . . . . . . . . . . . . 616 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 626.1 Apical View Classification and Localization . . . . . . . . . . . . 636.2 Segmentation and Volume Estimation of the LA . . . . . . . . . . 646.3 MR Diagnostic Labels . . . . . . . . . . . . . . . . . . . . . . . 716.4 Diastolic Dysfunction and LV Filling Pressure . . . . . . . . . . . 746.5 Time and Computational Complexity . . . . . . . . . . . . . . . . 777 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84viList of FiguresFigure 1.1 Overview of the proposed joint fusion information framework 6Figure 2.1 A diagram of the heart . . . . . . . . . . . . . . . . . . . . . 12Figure 2.2 Illustration of an echocardiography setup . . . . . . . . . . . 14Figure 2.3 Current protocol for echocardiography examinations . . . . . 15Figure 2.4 Example Left Atrial Segmentations . . . . . . . . . . . . . . 16Figure 2.5 Transthoracic Apical 4 Chamber View . . . . . . . . . . . . . 17Figure 2.6 Transthoracic Apical 2 Chamber View . . . . . . . . . . . . . 18Figure 2.7 A mitral regurgitation decision algorithm . . . . . . . . . . . 23Figure 2.8 A visual representation of the PISA velocity technique . . . . 24Figure 2.9 LV filling pressure algorithm . . . . . . . . . . . . . . . . . . 27Figure 3.1 Server workflow diagram . . . . . . . . . . . . . . . . . . . . 40Figure 3.2 Histogram of pixel spacing for each apical view . . . . . . . . 42Figure 3.3 Distribution of left atrial volume data . . . . . . . . . . . . . 45Figure 3.4 Example Apical 4 Chamber ultrasound with segmentation in-formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46Figure 4.1 Overview of the components of the framework . . . . . . . . 47Figure 4.2 Overview of the PCA preprocessing phase for the framework . 48Figure 4.3 Overview of the Apical View Classification process. . . . . . 49Figure 4.4 A schematic diagram of the LA segmentation, volume estima-tion and disease detection phases. . . . . . . . . . . . . . . . 50Figure 4.5 Schematic of the jICA method. . . . . . . . . . . . . . . . . . 52viiFigure 6.1 Initial centroid estimation results . . . . . . . . . . . . . . . . 63Figure 6.2 Refined centroid estimation results using an affine intensity-based registration. . . . . . . . . . . . . . . . . . . . . . . . 64Figure 6.3 Correlation coefficient results between estimated and manuallyobtained LA volumes . . . . . . . . . . . . . . . . . . . . . . 65Figure 6.4 A comparison of the DICE coefficient for both joint and singlesegmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . 66Figure 6.5 Example Segmentations . . . . . . . . . . . . . . . . . . . . 67Figure 6.6 MDL Criterion results for jICA modelling. . . . . . . . . . . 68Figure 6.7 LA segmentations visually grouped in columns based on theirfirst principal components . . . . . . . . . . . . . . . . . . . 69Figure 6.8 Visualize the separability of Apical 4 Chamber (AP4) joint In-dependent Component Analysis (ICA) reconstruction weights . 70Figure 6.9 An example of shadow artifact in the LA. . . . . . . . . . . . 71Figure 6.10 The confusion matrix for the healthy-moderate/severe and cor-responding metrics. . . . . . . . . . . . . . . . . . . . . . . . 72Figure 6.11 Learned AP4 and AP2 joint independent sources . . . . . . . 73Figure 6.12 The Receiver Operating Characteristics (ROC) curve for thehealthy-moderate/severe Support Vector Machine (SVM) clas-sifier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74Figure 6.13 Visualization of influence regarding normalized body surfacearea (a) and LA volume (b) on False Positives and True Nega-tives for MR classification. . . . . . . . . . . . . . . . . . . . 75Figure 6.14 The affect of normalized body surface area (a) and LA volume(b) patient features on MR classification’s False Positives andTrue Negatives. . . . . . . . . . . . . . . . . . . . . . . . . . 75Figure 6.15 Analysis of MR classification versus mitral annulus surface area. 76Figure 6.16 Diastolic dysfunction confusion matrix . . . . . . . . . . . . 77Figure 6.17 Feature selection using mRMR comparing classification accu-racy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78viiiGlossary2D Two-Dimensional2DE Two-Dimensional Echocardiography3D Three-Dimensional3DE Three-Dimensional EchocardiographyAAM Active Appearance ModelAAMM Active Appearance Motion ModelAICD Automatic Implantable Cardioverter DefibrillatorAP2 Apical 2 ChamberAP4 Apical 4 ChamberASE American Society of EchocardiographyASM Active Shape ModelBC British ColumbiaBSA Body Surface AreaCMR Cardiovascular Magnetic ResonanceCSV Comma-Separated ValuesixCT Computed TomographyDD Diastolic DysfunctionDICE Quantitative Dice IndexDICOM Digital Imaging and Communications in MedicineECHO EchocardiographyEF Ejection FractionEROA Effective Regurgitation Orifice AreaFN False NegativeFP False PositiveGUI Graphical User InterfaceICA Independent Component AnalysisIT Information TechnologyjICA Joint Independent Component AnalysisLA Left AtriumLV Left VentricleMAD Mean Absolute DistanceMDL Minimum Descriptive LengthMR Mitral RegurgitationMRI Magnetic Resonance ImagingmRMR Minimum Redundancy Maximum RelevancyMRN Medical Record NumberPACS Picture Archiving and Communication SystemxPCA Principle Component AnalysisPDM Point Distribution ModelPISA Proximal Isovelocity Surface AreaRBF Radial Basis FunctionRF Radio FrequencyROC Receiver Operating CharacteristicsROI Region of InterestSAX Parasternal long axisSSE Sum of Squared ErrorsSVM Support Vector MachineTEE Transesophageal EchocardiographyTN True NegativeTP True PositivetSNE T-Distributed Stochastic Neighbor EmbeddingTTE Transthoracic EchocardiographyUBC University of British ColumbiaUS UltrasoundVCH Vancouver Coastal HealthVGH Vancouver General HospitalxiAcknowledgmentsI offer my enduring gratitude to the faculty, staff and my fellow students at theUBC, who have inspired me to continue my work in this field. My deepest gratitudeis to my supervisor, Dr. Purang Abolmaesumi. I have been fortunate to have asupervisor who gave me the freedom to explore on my own, and at the same timethe guidance to recover when I faltered. Purang’s patience and support helped meovercome many crisis situations and finish this dissertation.I thank Drs. Teresa Tsang and John Jue for enlarging my vision of the fieldcardiology and helping me with my endless questions.A special thanks are owed to my parents. Thank you for supporting me through-out my years of education, both morally and financially.Finally, I would also like to thank my girlfriend, Adriana, and all of my friendswho have helped me stay sane through these difficult years.xiiChapter 1IntroductionCardiovascular disease is the number one cause of death globally [97]. Early di-agnosis of cardiovascular diseases is important for monitoring and treatment. Apatient’s cardiovascular risk is assessed by analyzing numerous parameters deter-mined from features in the imaging data such as endocardium boundaries, whichaids in measuring the Left Atrium and Left Ventricle volume. These measurementshave a direct impact on clinical management of patient outcomes.Cardiovascular diagnosis is performed using Cardiovascular Magnetic Reso-nance (CMR), Computed Tomography (CT), Three-Dimensional Echocardiography(3DE) and Two-Dimensional Echocardiography (2DE) imaging. While CMR and CTprovide high quality images of the heart, such methods are not routine: do not allowhigh-patient-throughput due to long acquisition times and limited availability; pa-tients with metal implants cannot be imaged using CMR; and patients imaged usingCT are exposed to ionizing radiation. Echocardiography (ECHO) is a non-invasive,low-cost, portable, and widely available imaging modality - making ECHO increas-ingly the standard for diagnosis of various cardiac conditions, risk stratification,and prognostication. While 3DE has several advantages over 2DE in terms of re-duced operator variability and improve ECHO workflow in terms of efficiency andaccuracy through automation, current 3DE technology has a lower spatial and tem-poral resolution vs. 2DE, 3DE images are difficult to interpret, and the majorityof sonographers, including at Vancouver General Hospital (VGH), only use 2DE inroutine clinical exams.1The most common 2DE interrogation is Transthoracic Echocardiography (TTE).TTE is used to assess the structure and functionality of the heart using both B-modeand colour Doppler flow imaging. In TTE the heart is imaged from at least six stan-dard views: parasternal long and short axes, Apical 2 Chamber, Apical 4 Cham-ber, subcostal, and suprasternal views. Despite the valuable role of TTE, analysisis operator dependent, subjective when visually detecting disease, and consider-ably variable in the quality of the measurements [14]. This variability comes frommultiple sources including sonographers’ experience, workload and time pressure.Reducing the variability of TTE analysis could lead to a more standardized measure-ment process of the patient to aid in preventing adverse cardiac events, includingheart failure or stroke [1, 91].ECHOs are obtained by the imaging technician, or sonographer, who is a med-ical professional accredited by the Canadian Medical Association to operate theUltrasound (US) machine. An example of a common challenge faced by sonog-raphers during 2DE acquisition is the foreshortening of the LV, which can causeinaccurate measurements by a cardiologist, who relies on optimal acquisition. An-other source of error stems from the ultrasound image formation, where variabilityof the probe’s angle not only makes measurement discrepancies, but also increasesthe difficulty of ECHO image segmentation too. The segmentation difficultly in-creases as the boundary between the chamber tissue and blood pool becomes lessdefined, which is related to:• Noise in inhomogeneous regions• Orientation of the view• Limited field of view through the patient’s rib cage• Irregular geometry of the LA• Maintaining orthogonal views between the AP4 and AP2.During a standard study, there are numerous parameters determined from ECHOdata, including endocardium boundaries, LA volume, Mitral Regurgitation (MR)severity and diastolic dysfunction. Several essential clinical measurements for di-agnosis are derived by delineating the LA’s boundary accurately in multiple TTE2views. Undefined boundaries, related to the points mentioned above, increase thevariability in measurements performed by 2DE and may have a direct impact onpatient care and the clinical management of heart conditions. Performing reli-able measurements, despite the inherent variability of 2DE, could lead to a morestandardized examination of the patient to aid in preventing irreversible cardiacdysfunction.One such essential clinical measurement is derived by delineating the LA’sboundary accurately in multiple ECHO views to correctly determine a patients max-imal LA volume. This volume is useful in assessment of risk for first ischemicstroke. Stroke is the third most common cause of death worldwide [60]. Clinicaltrials have suggested that assessment of LA volume provides a quantitative assess-ment of risk for first ischemic and an independent predictor of death [6]. Overesti-mation of LA volume may lead to device implantation of an Automatic ImplantableCardioverter Defibrillator (AICD) in patients who may not benefit from this expen-sive therapy. Underestimation may deprive some deserving patients of the survivalbenefit afforded by AICD. Furthermore, LA volume has also been used for investi-gational procedures to determine LA enlargement, since it has been discovered tobe a reliable marker for predicting stroke and mortality [35, 68, 80, 82, 107]. Anenlarged LA is also associated with severe MR, a condition that must be closelymonitored should it become symptomatic. Detecting MR before it becomes un-manageable allows physicians to alleviate diastolic dysfunction and safely manageheart failure [89], since without proper treatment, symptomatic patients have anannual death rate of over 5% a year [20].Determining LA enlargement can be performed using traditional 2DE or newer3DE methods. LA volumes determined from 3DE have been compared to CMR imag-ing and have demonstrated an improved accuracy over 2DE imaging with high cor-relation to CMR (r = 0.93) [69]. However, as of 2015, the American Association ofEchocardiography does not recommend the use of 3DE for assessing LA volume, asthere is limited data [5] and lack of standardized methodology to determine normalLA values [53].Currently, to determine LA enlargement health care professionals must segmentthe end-systole frame of LA in both apical views separately, which is a highly repet-itive process. Significant effort has been invested in standardizing LV ECHO cham-3ber segmentation through both semi-automatic and automatic analysis of cardiacdata over the last few decades, but it still remains an open problem due to thesementioned issues [72]. Many variations on segmenting the blood-tissue bound-ary have been proposed in the literature such as: morphological image process-ing methods [50], neural networks [13, 21], active appearance model [15], ac-tive shape models [29], convexity pursuit algorithms [19], shape regression ma-chines [40, 102] and active contour variants [10, 11, 39, 58, 64, 66, 83, 86]. Thesegmentation of the LA blood-tissue barrier has not been looked at extensivelywithin the literature when compared to the LV. Segmentation of the LA is con-sidered more difficult due to a) its more complex shape, b) there being less contrastbetween the endocardium and blood volume, c) the segmentation must cut throughthe LA appendage and through the mitral annulus, and d) the LA chamber is thefurthest away from the transducer; resulting in ECHO images missing importantanatomical features, such as the LA edges. However, the aforementioned LV seg-mentation methods do not deal with these obstacles, and do not incorporate theinformation from additional apical ECHO views.Additionally, other parameters could be improved by having an algorithm thataids cardiologists in edge cases where they are unable to classify patients basedon strict ECHO guidelines. During a standard study, cardiologists must mark apatient as indeterminate for the diastolic function or filling pressure, should theirpatient not meet the required parameters. As a result, the frequency of a patient’srecurring visits could be altered. Further incorporation of clinical ECHO parameterscould potentially be used to further define these indeterminate groups.Having a framework that would provide fast, consistent and accurate analy-sis of LA’s anatomy, volume, diastolic function and filling pressure would aid indiagnosing cardiovascular diseases, which would translate into direct benefits inpatient care. The demand for ECHO assessment is escalating given the aging popu-lation and the rapid increase in the numbers of patients with cardiovascular disease.Publicly funded resources are limited. Currently, at VGH and University of BritishColumbia (UBC) Hospital, the wait time for a non-urgent ECHO is 3 months. TheECHO laboratories across this province are facing the same crisis. There is a press-ing need for more efficient application of the automatic image analysis techniquesto meet this growing demand.41.1 Proposed FrameworkWe propose a joint information framework for fusion of ECHO image intensity in-formation and their segmentations from multiple 2D views of the heart to automat-ically estimate clinical parameters and diagnostic labels. Our proposed framework,using standard ECHO paired with clinical measurements, aims to reduce the numberof manual measurements performed during a standard ECHO study and to ease theclinical workflow and reduce measurement variability. Here, we introduce the JointIndependent Component Analysis model [18] for this framework to learn patternsfrom the observable correlation between each ultrasound intensity voxel and thecorresponding segmentation label. We use these patterns for LA volume estimationand classification of individuals with MR. The framework consists of four compo-nents, as seen in Figure 1.1: 1) model generation and alignment; 2) localization; 3)joint source reconstruction; and 4) classification and estimation of patient labels.We use the jICA framework to reduce our large database into compact, maximallyindependent basis functions, which combine intensity and shape information intoa unified space to reveal diagnosis labels. As a corollary objective, we analyzefinal common measurements performed during routine studies to predict diastolicdysfunction and filling pressure for patients.1.1.1 ContributionsOur research goal has been to develop a framework for LA volume estimation anddisease label analysis on ECHO imaging information. In the course of this project,the following contributions were made:• Obtaining a largest of its kind dataset of medical information by interfacingwith VGH’s medical ECHO storage system (Philip’s XceleraTM) and clinicaldatabase (FilemakerTM).• Proposing a LA segmentation technique that uses the joint information fromdifferent apical views.• Investigating a new technique for classifying normal and moderate or moresevere mitral regurgitation based on jICA reconstruction coefficients.5Alignment	and	Model	Generation• Apical	view	selection• Register	database	of	image• Stratify	patients	based	on	their	anatomyLeft	Atrium	Localization• Query	database	for	most	similar	image• Initialize	template	matching	algorithm	• Define	left	atrium	regionJoint	Source	Reconstruction• Generate	multiple	joint	ICA	models• Calculate	segmentation	probability	maps	• Perform	optimal	segmentation	selectionClassification	and	Estimation• Calculate	left	atrium	volume• Categorize	by	mitral	regurgitation	severity	Figure 1.1: Overview of the proposed joint fusion information framework.The main processes of the four steps are unpacked above.6• Investigating the use of combining clinical features for analysis diastolic dys-function and filling pressure to prevent indeterminate classification.1.2 Structure of ThesisThis thesis consists of seven chapters. We provide an overview of each chapter andthe associated contributions below:Chapter 2: Background and Related WorksIn this chapter, we review the basic anatomy of the heart to gain an un-derstanding of its global function and the importance of the LA within thepulmonary circuit. We also provide a detailed walk-through of a typicalsonographer routine study, specifically discussing patient positions, imageacquisition protocol, and various measurements for the cardiologist’s re-port. The work flow of the attending cardiologist and their responsibilitiesare also revealed. Furthermore, we demonstrate the importance of the LAby discussing LA dynamics, measurement protocols, clinical importance forsymptom management and disease related to LA dynamics, such as MR.Chapter 3: MaterialsIn this chapter we describe the datasets obtained from the Philip’s XceleraTMand FilemakerTM systems, which were used to train the multiple modelsemployed by the joint framework. The ECHO data materials were obtainedfrom XceleraTM, which were extracted from routine studies since as earlyas 2005 from multiple sonographers and ultrasound machines. We will usethis ECHO data in future chapters. Patients’ clinical measurements wereacquired from the FilemakerTM database; this data contains over 200,000records and were used to label patient ECHOs with their correct disease la-bels and volume information.7Chapter 4: MethodsWe explore:1. Localization: Automatic localization of the LA Region of Interest(ROI) from AP4 and AP2 ECHO views. We use this localization stepto create a bounding box around the anatomy, such that only the lo-cal variations of the LA are learned through jICA. To determine thebounding box that encompasses the entire LA region we use a cen-troid estimation matching technique paired with an intensity-basedregistration.2. Fusion Analysis for Volume Estimation: We propose to classify indi-viduals based on the anatomical shape of their LA chamber to create aset of structure-specific models. We aim to create a comprehensive setof models by leveraging our diverse database to stratify patients intocohorts. Patients are grouped based on their coefficients from Prin-ciple Component Analysis (PCA) decomposition of their AP4 and AP2segmentations. Here, we use the set of models to generate spatially in-dependent joint sources using jICA within a Support Vector Machineclassification method [32]. Each model’s joint sources are used toproduce a segmentation probability map through linear recombina-tion with a patient’s ECHO image. The reconstruction profiles fromeach model were used within an SVM component to select the modelmost capable of representing the anatomy’s structural variations withthe most representative reconstruction coefficients. Post-processingof the AP4 and AP2 segmentation labels are performed to calculate LAsurface area and perform volume estimation.3. Disease Label Estimation: We use both the feature information fromthe joint sources outputted from the jICA framework and common clin-ical measurements to create a disease prediction component. We firstlook to separate healthy patients and those with moderate MR solelybased on the information gathered from 2D M-mode. Furthermore,8we seek to improve the current standard-of-care diastolic dysfunc-tion classification and filling pressure to prevent indeterminate patientidentification. This goal is accomplished by using a continuous spec-trum of features for classification, instead of current discrete clinicalguidelines, using a multi-stage SVM component and feature reductionmethod, Minimum Redundancy Maximum Relevancy (mRMR) [76].Chapter 5: ExperimentsTo evaluate the presented joint segmentation and labeling framework, weperform a validation on 6,993 studies that were acquired from the routinecardiology care. We use the end-systolic frame of each acquired ECHO cine,to classify the the current apical view, clinical LA segmentation contour forAP4 and AP2 views, and MR disease labels. Additionally, using solely onlyclinical measurements, we assess diastolic dysfunction and filling pressurefrom the archival data.Chapter 6: Results & DiscussionWe present the results of the proposed segmentation method and the classi-fication of mitral regurgitation. Using the segmentation method, we inves-tigate and quantify the added advantage of the joint information model forsegmentation and MR classification over the single view information model.Next, we analyze the major modulation of these sources created by jICA forthe single and combined models and discuss their clinical interpretation. Asan example, we perform MR identification using the jICA sources to achieve82% identification accuracy between healthy patients and those with mod-erate and severe mitral regurgitation. We also perform diastolic dysfunctionand filling pressure classification of varying severity levels. Finally, we vi-sualize the ECHO ROI data by using T-Distributed Stochastic Neighbor Em-bedding (tSNE) [34] to learn how the reconstruction coefficients outputtedfrom the framework represent the variability in LA anatomy.9Chapter 7: Conclusion and Future WorkWe conclude the thesis with a short summary followed by the major contri-butions along with the suggestions of future work in this area. We also re-view potential improvements, like enhanced model selection for the frame-work, to increase segmentation and classification accuracy.10Chapter 2Background2.1 Anatomy of the HeartThe heart is a muscular organ that supplies blood through the circulatorysystem. It is located in the middle of the mediastinum, behind the breast-bone. As seen in Figure 2.1, the supplied de-oxygenated blood from the su-perior and inferior vena cavae enters the right atrium, which is then pumpedto the right ventricle. The right ventricle passes the blood to the pulmonarycircuit, where the blood becomes oxygenated and feeds into the LA. Insidethe LA, the blood is further pumped into the left ventricle, through the mitralvalve, and out through the aorta.The pumping of blood through the four chambers follows a cardiacrhythm. The rhythm of a heartbeat can be broken up into two components:systole and diastole. In systole, the ventricles in the heart contract and thenare followed by a relaxation, which occurs in diastole. Conversely, the atriaperform the opposite action to the ventricle within each phase. For example,when the LV begins to relax in diastole, the LA begins to contract. The LA isat its maximal volume during end-systolic, which is before the opening ofthe mitral valve. Blood is pumped most efficiently when the ventricles andatria work in concert.The heart’s wall is made up of three layers, the endocardium, the my-11Figure 2.1: A diagram of the heart. The white arrows demonstrate the di-rection of blood flow through the valves, arteries and veins. This im-age is available under a Creative Commons Attribution Licence 2.0 at and the epicardium. The innermost boundary that interfaces withthe blood is the endocardium. The endocardium is joined with the my-ocardium, the middle layer, which is the connective muscle within the heart.Finally, the outer layer of the heart is the epicardium that supplies bloodvessels and nerves.2.2 Ultrasound ImagesThe basic ultrasound image is formulated by first using an ultrasound trans-ducer to first transmit and then receive Radio Frequency (RF) signals. Thesesignals are converted to digital RF signal, which is then filtered to producean envelope-detected signal. The envelope-detected signal then undergoespost-processing to produce the final B-mode image. Additionally, ECHOclinics also interpret colour flow Doppler for measurements.122.2.1 B-ModeThe B-mode ECHO is the most common method of imaging for both 2DEand 3DE, which produces a visual interface for the interrogated anatomy.The position of the ECHO is determined from the transducer’s angle and thetransmit time of the ultrasound signal. This image acquisition method isreal time, allowing up to 50-70 images per second in 2DE.2.2.2 Colour DopplerColour Doppler allows medical specialists observe the blood flow betweenchambers of the heart. This is done by colour-encoding Doppler informa-tion and overlaying the colours of the 2D ECHO image. Each colour rep-resents the speed of the blood flow within the ROI. Blood flow towards thetransducer is red and flow away from the transducer is blue. These coloursalso vary with shades of red and blue depending on the velocity and di-rection. Colour Doppler can be used to view blood flow in many areasof the heart simultaneously; however, colour Doppler only allows semi-quantitative assessment of regurgitant blood velocity. The severity of theregurgitant jet is limited by technical and physiological variables that af-fect the appearance of the jet. Eccentric jets will travel along the wall ofthe atrium and will provide erroneous flow measurements, leading to mis-classified regurgitant severity.2.3 EchocardiographyIn a modern ECHO laboratory, a study’s workflow is broken up into twostages: a) examination stage performed by the sonographer, and b) the re-porting stage performed by the cardiologist (Figure 2.3). Diagnosis relies onaccurate measurement of cardiac parameters from a sonographer’s interpre-tation of ECHO views. However, due to the low-resolution nature of ECHOand subjective judgment of sonographers, many of these segmentations lackin precision [12] and suffer from large observer variability [38]. The cur-rent unidirectional process (sonographer! cardiologist) makes it difficult13Figure 2.2: A standard echocardiography setup. The patient lies on their sideas a sonographer moves the transducer against their chest. This im-age is available under a Creative Commons Attribution Licence 2.0 at improving accuracy of measurements in busy clinical laboratories. Be-low, we describe the workflow in the context of VGH ECHO Laboratory, asan example of a modern ECHO centre.2.3.1 Sonographer WorkflowAcquisitionThe sonographer’s workflow in a modern ECHO laboratory is time and labourintensive. Before the acquisition begins, a sonographer must spend timepositioning a patient to obtain the best views, which can be a cumber-some process with the elderly. Patients are usually positioned lying per-pendicular to the bed on their left side to allow for optimal imaging andergonomics (Figure 2.2). Next, in the acquisition phase the sonographer14Sonographer* Cardiologist*Exam%Stage:%• Image*Acquisi8on*• Ini8al*Measurements*• Ini8al*Report*Repor/ng%Stage:%• Review*Images*• Modify*measurements*• Create*final*report*Figure 2.3: Current protocol for echocardiography examinations, each stagerepresented by a different hospital staff member. This image isavailable under a Creative Commons Attribution Licence 2.0 at follow a standard protocol to ensure consistent transducer orientationsbetween imaging technicians. The views are obtained a sonographer plac-ing a transducer on the patient’s chest wall (Figure 2.2). The AP4 view isachieved by placing the transducer at the apex (Figure 2.5), which allowssimultaneous viewing of all four chambers. The AP2 (Figure 2.6) view isimaged counter-clockwise and perpendicularly to the AP4 imaging plane.Only the LV and LA can be witnessed in this plane. In the acquisition stage,the sonographer acquires electrocardiogram gated cine clips from a pre-defined list of views found in Table 2.1. For each view, the sonographerreviews each cine clip and chooses a section, containing single or multiplecardiac cycles, which best interrogates the desired anatomy. Additionally,if a sonographer identifies an emerging pathology during the acquisition,such as mitral regurgitation or a thrombus, the patient must be kept whilea cardiologist is consulted. Sonographers are considered the first line of15Figure 2.4: Example LA segmentations of the AP4 (left) and AP2 (right).The area is estimated in each image using the method of disks,where each view’s area is the summation the many disks. This im-age is available under a Creative Commons Attribution Licence 2.0 at 2.1: Acquisition view protocol for sonographers to obtain cine.View ProcedureParasternal Long Axis M-mode Sweep, 2D measurements, right ventricularinflow viewParasternal Short Axis 2D images, 2D measurementsApical 4 Chamber 2D images, tissue Doppler, 2D measurementsApical 5 Chamber 2D images, tissue DopplerApical 2 Chamber 2D images, tissue DopplerApical 3 Chamber 2D imagesSubcostal 2D imagesAortic Arch 2D imagesRight Parasternal 2D imagesdiagnosis and cardiologists rely upon the sonographers skill to observe anddiagnose any unforeseen pathology and avoid chamber-foreshortening. Ifsub-optimal images are selected, subsequent measurements and diagnosesmay be affected.16Figure 2.5: (Left) An Apical 4 Chamber transthoracic echocardiogram dis-playing the right ventricle (RV), left ventricle (LV), right atrium (RA)and right ventricle (RV). (Right) An anatomical diagram of the Apical4 Chamber view. This image is available under a Creative CommonsAttribution Licence 2.0 at Computer WorkflowSonographers are responsible for making many of the measurements on theviews obtained, seen in Table 2.2. Routinely, sonographers delineate thechambers in ECHO images, i.e. LV and LA (Figure 2.4) to measure car-diac parameters manually. Maximal LA volume is measured by tracing theblood-tissue boundary at end-ventricular systole. These measurements areperformed on the ECHO images, which are uploaded automatically from USworkstations to the ECHO storage system, XceleraTM (Philips Healthcare,Netherlands). Once the sonographer finishes collecting measurements, theyare entered into a preliminary report on VGH’s FilemakerTM database. Af-terwards, a cardiologist enter the final report using into their custom recordsdatabase, FilemakerTM (Subsidiary of Apple, California), after reviewingthe measurements and cine clips from XceleraTM.17Figure 2.6: (Left) An Apical 2 Chamber transthoracic echocardiogram dis-playing the left ventricle (LV) and left atrium (LA). (Right) Ananatomical diagram of the Apical 2 Chamber view. This imageis available under a Creative Commons Attribution Licence 2.0 at 2.2: Common structures assessed during sonographer acquisition.Structure MeasurementsMitral Valve Qualitative assessment of structure, function and re-gurgitationAortic Valve Qualitative assessment of structure, function andleafletsTricuspid Valve Qualitative assessment of structure, function and re-gurgitationPulmonary Valve Qualitative assessment of structure and regurgitationAorta Diameter and pericardial effusionLeft Atrium Function, diameter, volume, areaLeft Ventricle Function, filling pressure, wall movement and ejec-tion fractionRight Ventricle Structure, function and width182.3.3 Cardiologist WorkflowIn the reporting stage (Figure 2.3), a cardiologist reviews the sonographer’scine clips, comments, and confirms the correct landmarks were used forthe measurements. These measurements, along with the final diagnosis,are used to create the final report. The quality of the subsequent report aredirectly based on the availability of appropriate images and the accuracy ofmeasurements. If the physician is unsure that the ECHO views accuratelyreflect a patient’s pathology, a patient must be called back for a follow-upvisit. Generally, cardiologists will not interact with the patient.2.4 Echocardiography Diagnosis2.4.1 Left Atrium DynamicsThe left atrium is located on the left posterior side of the heart, and it isone of four heart chambers. The LA is also the furthest chamber from theTTE probe, making it difficult to image. The LA acts as a reservoir forblood returning from the pulmonary circuit, which is then pumped to theleft ventricle of the heart during ventricular systole (contraction). The LAcan be visualized via transthoracic or, the more invasive, transesophagealechocardiography. Two-Dimensional (2D) imaging of the atria can provideimportant information that aids a clinician’s diagnosis or can be used forguidance during surgery. Imaging of the LA has traditionally been used tomonitor for thromboembolic events, but has more recently been used morefor investigational procedures to determine LA enlargement, since it hasbeen discovered to be a reliable marker for predicting heart failure, strokeand mortality [35, 68, 80, 82, 107].During routine patient visits, assessment of the LA is included in stan-dard transthoracic 2DE imaging studies. From our findings, the LA is al-ways segmented in a typical patient to assess atrial enlargement, whereas LVsegmentation occurs only when requested by a clinician. A clinical study,aimed to determine how common LA enlargement was, found that in a ran-19domly sampled population of 2,000 patients, 16% of both sexes exhibitedLA enlargement [77]. It is common knowledge that as patients age, their LAchamber naturally increases in volume. Regardless, special attention is paidto the LA anatomy as changes in its morphology can be associated with anumber of disease states like diastolic dysfunction, mitral stenosis, MR, andatrial fibrillation.Commonmeasurements to discern the size on the LA chamber in 2DE area) comparing the diameter of the LA to the diameter of the aorta, where LAenlargement is compared to aortic dilation; b) 2D single plane volume esti-mation; and c) 2D biplane volume estimation. The most performed methodin clinical practice to determine left atrial size is 2D biplane volume esti-mation, as this method requires the least geometric assumptions and incor-porates the irregular shape of the LA [54]. Calculating LA volume via 2DEimaging requires two orthogonal apical views, the AP4 and AP2 (Figure 2.4).Volume estimates can be calculated using the area-length algorithm as wellas the method of disks; both methods have been validated using angiogra-phy and cardiac CT [49, 84]. The direct atrium volume estimation (VLA) iscalculated by analyzing the segmentation labels from each orthogonal viewto determine the apical chamber areas (A4C,A2C) and the atrium’s length(L4C,L2C). The area can be calculated by summing the pixels within thesegmented regions or by using planimetry. The planimetry technique firstdivides each apical area into at least 20 disks with a variable width (ai) andfixed height (bi), and then sums their individuals areas together (Figure 2.4).Once the area has been calculated, the volume can be estimated using themethod of disks (Equation 2.1), single plane area-length (Equation 2.2) orthe preferred area-length 2D method (Equation 2.3) [53]:VLA =p420Âi=1ai⇥bi⇥ max(L4C,L2C)20 (2.1)VLA =83p⇥ A2L(2.2)20Table 2.3: Recommended chamber quantification for LA volumes.Severity Male (mL) Female (mL) Indexed (mL/m2)Normal 18 - 58 22 - 52 22 ± 6Mild 59 - 68 53 - 62 28 - 33Moderate 69 - 78 63 - 72 34 - 40Severe  79  73  40VLA =83p⇥ (A4C⇥A2C)min(L4C,L2C). (2.3)Clinically, the LA long-axis dimension is defined as the distance per-pendicular from the center of the mitral valve annulus to the apex of theatrium. However, in practice, the major axis is approximately perpendicu-lar to account for unusual morphology, like LA enlargement. Additionally,determining the boundary of the LA to calculate the area for the above for-mulas can be increasingly difficult in sub-optimal images, as the boundariesof the endocardium’s soft tissue are not well defined.These volumes are then indexed with the patient’s Body Surface Area(BSA) using the clinically standard Du Bois formula (BSA = 0.007184⇥Weight0.425⇥Height0.725). This formula is used to normalize the LA sizebetween men and women and removes the gender difference in LA size [51].This allows the indexed LA volume to be compared against a normal popula-tion’s volume measurements (found in Table 2.3), which allows a patient’satrial function and pathology to be stratified. These patients can then beplaced into categories that are associated with their risk and treatment op-tions [53]. There have been many indexing methods proposed; however,BSA indexing is the only method recommended by the American Society ofEchocardiography [53]. Stratification guidelines have been recommendedby the American Society of Echocardiography that outlines 28 mL/m2 asthe maximum normal LA indexed volume and severe dilation occurs at avolume of  40 mL/m2 [52].212.4.2 Mitral RegurgitationThe structural phenomenon, MR, is caused when the heart’s mitral valvedoes not close properly during a high-energy transfer of blood. The high-energy blood flow gradient is between the LA to the LV. Symptoms of thedisease are caused by the incomplete seal of the mitral valve, a valve thatis intended to prevent blood from flowing backward into the LA cham-ber, effectively decreasing the efficiency of the heart. Mitral regurgitationis recognized as one of the most common valvular heart diseases, whichcan require surgical intervention through surgical or percutaneous meth-ods [96]. The early evaluation and detection of MR can be assessed semi-quantitatively by colour Doppler ECHO [71] and can be performed usingeither transthoracic or transosophogeal methods. Color flow imaging canreveal a jet of blood that enters the LA during ventricular systole. It is com-mon practice to judge the regurgitant jet based on its size relative to theLA anatomy. A jet that occupies  20% is said to be mild and a jet thatcovers more than 40% is said to be severe MR. Visual diagnosis of MR,without colour flow Doppler, is more difficult because abnormalities asso-ciated with MR are not easily identifiable using ECHO (unless there is cleardamage to the mitral valve). Instead, colour flow Doppler is more sensitiveto detect MR severity, as opposed to focusing on the structure of the cham-ber and valves. Colour flow Doppler ECHO is used to examine the velocityof the regurgitant blood flow during the diastole phase. The measurementsobtained from Doppler are used to calculate the transmitral waveform andthe peak velocity across the mitral valve. Without these measurements, dis-tinction between intermediate grades of MR severity is difficult, specificallybetween mild MR and moderate MR. Severe MR is most easily identified inpatient’s due to severe LA enlargement. Additionally, CMR can be essentialin evaluating eccentric MR, where the regurgitant jet’s measurements cannotbe calculated from standard ECHO.VGH sonographers stratify patients using the decision tree similar toFigure 2.7 based on the following visual estimates. Initially, patients are22Clinical'evalua+on'and'Doppler'echo'Visually,'is'regurgitant'jet'length'and'area'abnormal?'Yes:'Perform'PISA,'vena'contracta'diameter,'jet'length'and'area'measurements.'Moderate'mitral'regurgita+on'or'greater.'No:'Record'observa+ons.' Normal'func+on'or'mild'regurgita+on.'Figure 2.7: An algorithm that dictates a sonographer’s workflow at VGH. Thisdecision tree indicates whether further measurements should be per-formed to quantify the severity of MR.screened based on the relative length, area and colour of the regurgitant jetfrom the colour flow Doppler ECHO. However, the size of the jet can be mis-leading to a sonographer, as it takes an experienced sonographer to properlycalibrate the equipment to correctly show the size and color of a regurgitantjet [92]. If the jet is of moderate severity or more, further measurements,such as measuring the Proximal Isovelocity Surface Area (PISA) to calcu-late the Effective Regurgitation Orifice Area (EROA), are then performedand compared against the guidelines for severe MR. Essentially, the PISAsignal creates a layered hemisphere that expands out of the mitral valve intothe ventricle (Figure 2.8). Each layer is colour coded corresponding to thevelocity of the jet, and as the hemispheric shells get smaller in area, thevelocity within each shell increases. The PISA signal is used to measure theestimated radius of the ring closest to the mitral valve (r), the ring’s velocity(VPISA) and the peak velocity of the jet (VPk) to calculate the EROA [92]:EROA=(2pr2⇥VPISA)VPk(2.4)The calculated EROA is the most predominate method used to quantifyMR severity due to extensive testing [92], but it does have numerous limita-tions. These limitations are, PISA measurement makes geometrical assump-tions, calculating the PISA measurement requires an experienced sonogra-23RegurgitantRadiusMitral	Valve	JunctionMitralLeafletNormal	Flow	DirectionLA	ChamberLV	ChamberMitralLeafletFigure 2.8: A visual representation of the PISA technique, where flow is aseries of concentric hemispheres that decrease in size and increase invelocity.pher, and assessment of multiple and eccentric jets is not feasible. Whenexperienced sonographers perform PISA measurements correctly, the mea-surements are considered to be accurate [108]. Special attention must bepaid to eccentric regurgitant jets since they can be difficult to spot. If thejet is missed, there is an opportunity for a misdiagnosis by the attendingcardiologist.Determining a patient’s MR severity is not only important to stratifypatients, but it also dictates the frequency of serial visits specified by the2014 American College of Cardiology/American Heart Association guide-lines. These guidelines recommend that patient’s with no or mild MR arere-evaluated every 3-5 years, patient’s with moderate MR should be testedevery 1-2 years and those with severe MR should be seen every 6-12 monthsor sooner [71]. The goal of serial monitoring allows frequent assessmentof changes to a patient’s LV function and helps prevent irreversible damageeither by corrective surgery or drug treatment.In the literature, there are semi- and fully-automated methods capa-24ble of modeling the shape and motion of the mitral annulus from Three-Dimensional (3D) Transesophageal Echocardiography (TEE) and CT images.Existing methods to analyse the mitral annulus’s shape and motion includesemi-automated [85, 93] or fully automated [45, 62] valve modeling meth-ods. However, these methods either employ 3D TEE or CT images that areof a higher quality when compared to 3D TTE. Two mitral annulus modelingmethods for TTE have been proposed to provide quantitative measurementsrelated to MR. Grady et al. presented the first system to measure PISA ona 3D TTE ultrasound by segmenting Color Doppler volumes to successfullyclassify patients into mild-moderate and moderate-severe regurgitation cat-egories [41]. More recently, Wang et al. proposed a new framework toautomatically quantify MR jet volume and EROA against expert measure-ments [96]. Using 3D features of both B-Mode and Colour Doppler, theycreated a classifier to detect MR jet position with good results. Recent ad-vances in 3DE have made it possible to acquire volumetric and hemodynam-ical data simultaneously, which could aid in improving the accuracy of MRmeasurements.2.4.3 Diastolic FunctionHeart failure, a common and potentially fatal disease [63], can be caused bytwo major mechanisms: systolic dysfunction and diastolic dysfunction. Di-astolic heart failure occurs when a patient experiences signs of heart failure,but exhibits normal Ejection Fraction (EF), normal maximal LA volumesand symptoms of diastolic dysfunction. Nearly one-half of congestive heartfailures are categorized as diastolic heart failures [78], or associated withit [104]. The minimum diagnostic criteria for diastolic heart failure is out-lined by the Heart Failure Society of America and it recommends that: a)there is clinical evidence of heart failure; b) by definition, normal EF levels;c) LA enlargement; and d) evidence of diastolic dysfunction [59]. Gener-ally, Doppler ECHO is used to determine the severity of diastolic dysfunctionand aids in the diagnosis of diastolic heart failure, but diastolic dysfunctiondoes not indicate heart failure [106]. In particular, Doppler ECHO is used to25investigate the relaxation properties of the myocardium and stiffness of theLV since diastolic dysfunction is characterized by an abnormal relaxation,stiffness of the cardiac walls [105].ECHO indicators sought for diagnosis of diastolic dysfunction includeLA volume, transmitral Doppler inflow velocity patterns, pulmonary venousDoppler flow patterns, tissue Doppler velocities and M-mode flow propa-gation velocity. These measurements are then used in a decision tree tograde diastolic function from mild to severe [70]. There is also a trend thatgenerally shows the severity of diastolic dysfunction also increases with apatient’s age, which can increase their risk for a heart failure event [46].However, it has been indicated that these guidelines proposed are too strictto fully define certain patient groups with abnormal diastolic dysfunction,and according to the guidelines, be labelled as having indeterminate dias-tolic function [26]. This label presents a problem to clinicians who wish toreach a consensus on diagnosing diastolic heart failure.Filling PressureIn the ventricular diastole phase phase, the blood pressure in left and rightventricles will drop. As the pressure in the LV decreases, it will reach atrigger point where the LV pressure is less than the pressure in the LA. Con-sequently, the mitral valve will open, like a trap door, filing the LV withblood that was previously pumped into the LA. Likewise, the process alsooccurs in the right side of the heart, allowing blood to circulate in the heart.In both sides of the heart, the transfer of blood to the ventricles occurs intwo steps. First, after the mitral valve opens, blood will rush in at a fillingvelocity, E. Next, as some blood will remain in the atrium, the atrium willcontract to push the remaining blood volume into the ventricle. This fill-ing speed is labelled as the A filling velocity. Together, the E/A ratio is ametric to compare the rate of early to late ventricular filling velocity. In ahealthy heart, the E filling velocity should be greater than the A velocity.With disease and aging, the E velocity will slow, lowering the E/A ratio.Additionally, the velocity of the blood through the mitral annulus is called26the annular velocity (E’). The ratio of E/E’ is also another metric used fordiagnosis.When grading a patient’s diastolic dysfunction severity, clinicians willuse the LV filling pressure as an indicator. The patient’s LV filling pressureis determined using an algorithm, similar to Figure 2.9. Note, that thereexists opportunities for LV filling pressure to be subjective when the fillingvelocities do not match the decision tree’s cutoff values.Left	ventricle	ejection	fraction	is	normalNo,	estimate	mitral	inflow	velocities	(E	&	A)E/A	filling	velocities >	1 Normal LV	filling	pressureE/A	filling	velocities <	1	and	dilated	left	atriumLikely	restrictive	filling,	confirm	with	additional	Doppler	findingsYes,	estimate	mitral	inflow (E)	and	mitral	annuluar (E’)	velocityE/E’	<	8		 Normal	LV filling	pressureE/E’	>	13	 Elevated	LV	filling	pressure8	< E/E’	< 13Additional	Doppler	findings	are	required	to	determine	filling	ratingFigure 2.9: Algorithm for estimating LV filling pressure.2.5 Related Segmentation WorksSegmentation of ultrasound images is one of the most common image pro-cessing tasks in the biomedical field, but still remains an open problem dueto the variability of ultrasound image quality. Traditionally, ultrasound im-ages have been segmented using B-mode information and ultrasound RF sig-nals, with the former method being more conventional. We focus our back-ground search on the principle works of automatic segmentation of the en-docardium for echocardiography. In the literature, US imaging includes 2D,2D + t, 3D and 3D + t, although we focus on the 2DE and 2DE+t literature as it27is the most comprehensive and relevant to this work. In 2DE, great attentionhas been given to the implementation of automatic 2D LV endocardium seg-mentation, as it provides essential measurements for diagnosis. However,to the best of our knowledge, there is limited literature on LA endocardiumsegmentation since it has not been looked as extensively when comparedwith the LV chamber [19]. In consideration of the limited proposed semi-or completely automatic left atrial segmentation methods, we will providean overview of closely related LV segmentation algorithms. Most LV seg-mentation methods have been performed using 2DE with various views, likethe Parasternal long axis (SAX), AP4 and AP2. While each individual viewhas its own unique image acquisition challenges, segmentation still remainsa challenging task due to the variability of anatomy positioning, abnormaland diseased pathology, and low signal-to-noise ratio [72]. However, therehas been no proposed segmentation approaches to overcome these obsta-cles by simultaneously incorporating the intensity information from boththe AP4 and AP2 ECHO to segment both LA views concurrently.Below, we outline of the related works for segmentation methods for theendocardium of the LV. We review several segmentation techniques, whichcan be divided into three main categories: (1) boundary-driven and region-based techniques, (2) model fitting techniques, and (3) RF techniques. Inthe case of multiple techniques being combined, they are grouped by theirmost related section.2.5.1 Boundary-Driven & Region-Based Segmentation TechniquesActive contour models (snakes), first investigated by Kass et al., is oneof the most common boundary-driven segmentation techniques used. Theparametric active contour method evolves a curve within an image domainand influences it to deform around a shape. The deformation of an ac-tive contour model occurs by imposing internal and external forces on thecurve. Where the internal force acts as a smoothness constraint and theexternal force advances the contour to a shape’s edge. Generally, activecontour models optimize these forces in an energy minimization approach,28but models can also add additional energy terms to their objective func-tion. These energy terms leverage the intensity statistics of ECHO imagesto segment contours within the cardiac chambers. There have been manyapproaches to propose external energy functions to find an optimal contour,examples include using optical flow to segment the LV [65], gradient vectorflow [99], and balloon models [27]. Furthermore, Mishra et al. proposedto segment the LV frame-by-frame, by using the contour from the previousframe to initialize the active contour in the next frame [66]. The activecontour function employed, used an additional energy term to describe thenon-linear mapping of the intensity gradient. Alternatively, Mignotte andMeunier used a shifted Rayleigh distribution to create an external energyterm that modelled the LV’s gray level statistics in Parasternal long axisimage [64]. The authors suggested using the Rayleigh distribution to over-come ECHO noise instead of intensity gradient features, since Wagner et al.had shown fully developed speckle follows the Rayleigh statistical distribu-tion model. As an alternative, de Alexandria et al.’s active contour approachconsisted of representing the contour’s external energy term by modelling aregion’s intensity information in a 1D Hilbert transform [33]. In doing so,they proposed a radial active contour technique, pSnakes for LV segmen-tation. Their promising method leveraged the fact that RF beams divergefrom a single point on the probe, enabling the use of polar coordinates torepresent intensity and filter out image noise.Another modification of the active contour is the geometric active con-tour method [23]. This is based on the level-set method [73] and curveevolution theory, which allows a curve to evolve or split based on topolog-ical changes. In this method, the evolution of a curve is independent ofparametrization, and can be represented as a level-set function. The level-set function is a computationally costly function that can evolve a boundarywith topological changes based on region-based intensities or edge-basedfeatures [73]. Level-sets are considered as an alternative to active con-tours for ECHO segmentation, and Yan and Zhuang was the first to con-sider using level-sets for LV segmentation by applying the traditional fast29marching algorithm [100]. Yan and Zhuang improved the algorithm by re-placing the level-set speed term [100]. This term was originally based onlocal image gradient, but was modified to incorporate average energy of thewhole curve. This was done to protect the speed term from being influencedby noise. Yan and Zhuang investigated an alternative level-set framework,which combined intensity gradient edge constraints and a region intensitydistribution term to automatically delineate a closed boundary curve [100].Their method was applied to 2D slices of 3DE data with good results. Lever-aging an image’s intensity distribution was also explored by Sarti et al., whoalso incorporated intensity ECHO information into a level-set formulation.Conversely, Sarti et al. used the Rayleigh model of speckle, instead of theintensity gradient for the prior feature [83]. This choice in prior by Sartiet al. to incorporate the statistical distribution of gray levels was shown toaid in overcoming the known ECHO low signal/noise ratio for LV segmen-tation. Dydenko et al.’s level-set method also assumed a Rayleigh inten-sity distribution, which performed segmentation and tracking via a level-setfunction that was constrained by gray level image statistics and a shapeprior [39].Overall, variants of active contours and level-sets can be grouped to-gether as deformable models. Deformable models have the advantage of notrequiring a training step, being able to easily alter a model with additionalenergy terms in their objective function and incorporate shape priors con-straints (seen below). However, deformable models are at a disadvantagewhen medical images are noisy, the endocardium boundary has protrusionsand the required initialization is not close to the region of interest. There-fore, the increasingly more common strategy is to combine prior knowledgeof the shape of the object and intensity descriptor for segmentation.2.5.2 Model Fitting TechniquesUsing statistical priors, based on a learned shape or the geometrical assump-tions of the heart, help overcome analyzing ECHOs with excessive noiseor missing tissue. The statistical shape modelling techniques that include:30shape priors, Active Shape Model (ASM), Active Appearance Model (AAM)or joint ASM/AAM models can be categorized by their shape prior restric-tions.Previous literature on shape prior restriction was extensively studied asit was shown to improve the accuracy and reliability of ECHO segmenta-tions [37]. Using geometric models, a chamber’s anatomy can be repre-sented through a set of parametric equations, requiring few tuning parame-ters and did not require training. However, geometric models are not wellsuited for complex shapes because as the complexity increases, so does thecomputational cost to align the model. Using a geometrical model, Hamouand El-Sakka first segmented the LV in a 2D B-mode ECHO using an exter-nal energy gradient vector flow snake [43]. The snake was trained usingtwo, third order hyperbolas to better represent the ventricle in systole anddiastole. To aid intensity or edge-based functions that failed on imperfectECHOs, several authors further exploited the LV’s shape as a truncated ellipseto create a prior model of the expected ventricle shape. Alessandrini et al.addressed ECHO segmentation with a geometrically constrained model, butfor myocardium tissue in the SAX view [2]. Alessandrini et al. using alevel-set framework model that consisted of two ellipses represented by 10parameters to perform segmentation [2]. Recently, Dietenbeck et al. alsoembedded a geometrical shape prior into their proposed level-set frame-work for AP4 segmentation [37]. They approximated the myocardium us-ing a geometric model with two hyperquadics, allowing asymmetric shapemodeling in any view. The shape prior was then combined with a thicknessterm that allowed for joint segmentation of endocardium and epicardial bor-ders.Another way to build shape restricted priors is to used a learned shapemodel from annotated a priori LV data, allowing a target image to be matchedwith its corresponding model. This was previously performed using eithera mean contour curve [24, 25] or probabilistic maps [81]. Chen et al. cre-ated a mean contour curve shape prior by utilizing the LV contour of anannotated database to segment the epicardial and endocaridal borders [24].31This method utilized the distance between the active contour and the priormodel to create a new energy term for their level-set framework. Chen et al.extended their prior model to incorporate an intensity profile with the meanannotated LV curve [25]. This allowed the intensity profile along multiplesegmentations to be compared to the prior intensity profile, which was usedto select the best segmentation.Cootes and Taylor developed an alternative technique to describe shapepriors, using PCA on parametric contours of face outlines to obtain the mainaxis of variation [28]. This technique became a common method to obtaina shape prior from a database by capturing the main shape components ofmanually contoured images, while discarding redundant information. Lev-enton et al. applied this PCA technique to 2D to LV ECHO data, but used thesigned distance function of the image contours to represent the referenceshape, instead of parametric geometric contours. Additionally, the sameshape prior representation proposed by Leventon et al. [56] was adoptedby Tsai et al. to calculate the parameters needed to minimize their energyfunctional for segmentation of the 2D LV data [90].Cootes et al. also introduced the Point Distribution Model (PDM), an al-ternative way to create a shape constrained prior learned from a trainingset of manually drawn contours [29]. Cootes et al.’s proposed shape priorwas a combination of a mean shape of the contour points and, using PCA,a model of the main modes of shape variation within the dataset. The PDMhas become a standard in medical image segmentation, where shape priorsaid in interpreting noisy images. The development of the PDM led to theASM, which Cootes et al. used to represent the variation of manual LV con-tours [29]. The ASM is a statistical model of the shape of an object, which isconstrained by the PDM to only vary in ways witnessed within the trainingdatabase. This prior shape knowledge was represented within a probabilis-tic frameworks, such as a Gaussian distribution [29] or mixture model [28].While using ASM priors in endocardial segmentation is advantageous dueto its ability to handle intensity variations [29], ASM requires a point-to-point correspondence when finding parameter values that properly fit the32model to a new image. Additionally, a level-set variant of the ASM wasintroduced by Rousson and Paragios [81] to avoid a Gaussian distributedprior shape, which was applied to the LV data by Paragios and Deriche [74].In an attempt to improve the ASM model, an extension to include intensityprofile information was investigated to develop the AAM [30]. The AAMshares similarities to the ASM mode, but includes the intensity variations ofthe image too. By combining both shape and intensity data, Cootes et al.introduced a new method of matching statistical models of appearance toimages, which also models appearance texture [30]. The AAM model wasadditionally improved upon by Bosch et al. by adding temporal information,using PCA to learn shape prior of ECHO LV segmentations for each positionwithin a cine of images [15]. Furthermore, Bosch et al.’s AAM + motionmodel did not assume standard Guassian distribution of image intensity,and instead used non-linear intensity normalization to match the intensitydistribution of the images. The normalization proved to be beneficial to theLV segmentation, qualitatively increasing the quality of their endocardiumsegmentation. A limitation of the AAM method is new clinical images areassumed to have similar appearance and tissue property variations to thosethe model trained upon. This led to other extensions to include motioninformation [42] and an AAM plus temporal information (Active Appear-ance Motion Model (AAMM)) for 3D segmentation [67]. Mitchell et al.’sAAMM was capable of segmenting ECHO temporal image sequences [67].Previously, Zhou [102] questioned if the characterization of the LV’s endo-cardium can be accurately represented by a linear model, such as the linearappearance model in Cootes et al.’s AAM. Non-linear methods have beenproposed, citing the need from variability of training sets derived from dif-ferent patients, sonographers, and ultrasound machines. Furthermore, theliterature suggests that the ASM proposed by Cootes et al. [29] is effective,but lacks a good one-to-one point feature criteria [102]. Zhou proposed anon-linear LV segmentation machine called a shape regression machine toovercome the mentioned issues [102]. Their proposed shape regression ap-proach was shown to overcome blurred boundaries in B-mode ECHO and33could handle missing LV boundaries too. Their shape regression machineused statistics of the shape, appearance, and anatomy to construct a model.2.5.3 Radio Frequency TechniquesThe above referenced proposed information approaches are based on theanalysis of B-mode ECHO images; however, some authors [10, 11] pro-posed using RF, a potentially more informative source than the envelope ofthe ECHO image. Bernard et al. formulated the segmentation problem us-ing the generalized Gaussian distribution statistics of the RF signal, whichwas used within a Maximum Likelihood framework to delineate the my-ocardium [10]. Bernard et al. extended this work by showing that RF canalso reliably describe a chamber’s blood pool and tissue area via a Gen-eralized Gaussian distribution [8, 9]. However, the assumption that imageintensities represented by RF can be modelled by a Gaussian distributionhas not been validated against a range of intensities found in clinical B-mode data. As such, it has been shown that the gamma distribution canbetter describe the envelope of the RF in several studies [16, 87]. Tao et al.compared the the validity of RF representation by gamma distribution bycomparing it against Weibull, normal, and log-normal distributions on car-diac images [87]. Recently, Bui et al. used local gamma distributions in thedata term of a level-set energy function to perform segmentation within 2Dand 3D simulated ultrasound images, outperforming previous local Gaus-sian methods [17].Validation of LiteratureRegarding the validation of the segmentation techniques reviewed, we havesummarized the main contributions for LV endocardial segmentation tech-niques. We present this information in Table 2.4 along with their modality,ROI, view, size of the validation dataset and validation results. Most meth-ods were evaluated by their Mean Absolute Distance (MAD) between theexpert segmentation results and their methods’ contours. However, some34papers have provided global errors of EF or LV mass error. Point to surfacedistance or MAD are usually within the range of 1-3mm for 2DE/3DE.2.6 Commercial ProductsCurrently, both semi-automatic and fully automatic 2D and 3D cardiac mea-surements have been commercially explored by Philips, Siemens, GEHealth-care, TomTec, DiACardio and TeraRecon. The most prominent cardiacsuites are Philip’s QLabTM (Philips Healthcare, Netherlands) and Siemen’sSyngoTM (Siemen’s Healthcare, Germany), TomTec’s Image-AreaTM plat-forms (TomTec, Germany) and GE Healthcare’s CardiacIQ SuiteTM (GEHealthcare, United Kingdom), with their features listed below in Table 2.5.All suites provide automatic methods for quantifying a wide suite of param-eters determined from ECHO, with the goal to increase patient throughput,while improving accuracy and workflow efficiency. Currently, VGH cardi-ologists use a Philip’s QlabTM plugin to manually perform their measure-ments. British Columbia (BC) Children’s Hospital performs their measure-ments using Siemen’s SyngoTM base platform for manual delineation oftheir 2D data.35Table 2.4: Validation of 2D ECHO image sequence on the LV endocardium.Modality Reference Year View Number of Patients Validation Results3D Wolf et al. [98] 2002 Transophogeal 20 LV Endocardial error: 3.4 ± 2.3 mm2D Bosch et al. [15] 2002 AP4 129 LV Endocardial error: 3.54 ± 1.62 mm2D Lin et al. [58] 2003 Long-Axis 24 LV Endocardial error: 1.64 ± 0.5 mm2D Sarti et al. [83] 2005 AP4 15 LV Endocardial error: 1.6 ± 1.8 mm2D Georgescu et al. [40] 2005 AP4 206 N/A3D Angelini et al. [3] 2005 Long Axis 10 Absolute Error for EF: 4.6%2D Yue and Tagare [101] 2008 Short-Axis (Phantom) 1 N/A2D Zhou [102] 2010 AP4 527 LV Endocardial error: 2.2 pixels3D Zhu et al. [103] 2010 Long Axis View 11 (Canine) LV Endocardial error: 1.4 ±3D Leung et al. [55] 2010 Short-Axis 35 LV Endocardial error: 1.19 ± 0.472D Dietenbeck et al. [37] 2012 AP4 & AP2 20 AP4 Dice: 0.93 AP2 Dice: 0.892D Carneiro et al. [22] 2012 AP4 12 N/A2D Cao et al. [19] 2014 AP4 10 LV DICE: 0.7136Table 2.5: Commercial cardiac suite feature comparison.Feature QlabTM SyngoTM Image-ArenaTM CardiacIQTMAuto 2D LV border detectionp p p pAuto 2D LA border detectionpAuto 2D right ventricle border detectionAuto 2D Motion Strain & Stressp p p pAuto 3D LV EFp p pAuto 3D right ventricle volumepAuto 3D Motion Strain & Stressp p pVendor Independentp37Chapter 3MaterialsThe proposed framework for automatic analysis of the left atrium usesthe datasets described below. Software applications interfacing with thesedatabases are also briefly described.3.1 EthicsTo evaluate the 6,993 studies that were acquired from routine cardiologycare, we went through an ethics approval process from Clinical MedicalResearch Ethics Board of Vancouver Coastal Health (VCH) (H13-02370)and consultation from the VCH Information Privacy Office. Throughoutthe analysis, a rigorous process of data anonymization, de-identificationand data encryption was followed based on the guidelines recommendedby the VCH Privacy Office. All of the ECHO and patient information (age,sex, height, weight and related health issues) are assigned a non-identifyingalpha-numeric code that ensures the risk of re-identification of participantsfrom the acquired data is low. The data was encrypted and password pro-tected using TrueCryptTM (TrueCrypt.Org, Czech Republic), a softwarethat has been independent audited that concluded no significant flaws werepresent1.1htt ps : // Phase II NCC OCAP f inal.pd f383.2 Echocardiography DataFor the purposes of this thesis, a large database that reflected a wide rangeof patient morphology and pathology was required. By interfacing withVGH’s Cardiology clinic, more than 7,000 ECHO studies were acquired.Each study contained a patient’s ultrasounds that were requested by a physi-cian. All obtained ultrasound ECHO images were from retrospective stud-ies performed at VGH. To access these ultrasound studies, the cardiologydepartment’s XceleraTM database was interfaced, allowing information tobe downloaded pertaining to patient follow-ups, emergency, and investi-gational ECHO studies. At VGH, ECHO information are primarily acquiredfrom Philips iE33TM or the portable GE Vivid qTM ultrasound machines.The Digital Imaging and Communications in Medicine (DICOM) studies ob-tained from these devices were cached and uploaded to VGH’s cardiologydepartment’s XceleraTM server. Once the ECHO information was uploaded,the information was then accessible via an XceleraTM workstation terminal.3.2.1 Echo Lab SoftwareThe XceleraTM ultrasound software allows both cardiologists and sonogra-phers to access all saved ECHO studies within a Graphical User Interface(GUI) interface. The software also acts as a server that accepts DICOMimages from Philips US machines and stores it. The XceleraTM softwarefunctions by acting as a traditional DICOM viewer for medical staff, plusit incorporates advanced features for sonographers and cardiologists. Forinstance, within the software suite there is an image measurement modulethat allows cardiac measurements to be drawn directly on medical imagesand saved for future examination. The measurement module is manual andhas no automatic segmentation features for LV and LA segmentation. Sono-graphers use the measurement package to measure a LA’s area and majoraxis in both the AP4 and AP2 view. These measurements are then manuallyinput into the FilemakerTM database, which then calculates the patient’s LAvolume.39Ultrasound Machine DataPatient InformationPACS Echo ServerFilemakerServerXcelera ServerCardiologist assessment of study performed in XceleraFigure 3.1: A diagram designed to show the relation between the Echo, Xcel-era and Filemaker databases in the context of a routine cardiology study.The XceleraTM software operates by maintaining two separate databases.It downloads the ECHO information from the ultrasoundmachines and writesthis information to two of its own mySQL instances, echo and xcelera.XceleraTM’s echo database stores all the manual segmentation informationmade on an image, and xcelera stores patient information keys, which al-lows studies to be reopened with the correct measurements attached. Mean-ing, whenever measurements are performed within the software suite, themeasurements are saved to the echo database with a unique identifier. Thisidentifier key is also stored on in the xcelera database, which has an addi-tional identifier to link each particular ultrasound. The two databases worktogether using a key matching method to store large amount of data. Forthe purposes of this thesis, we only use the XceleraTM database to obtainECHOs, segmentation labels and major axis coordinates. XceleraTM (PhilipsHealthcare, Netherlands)’s database relationship chart can be found in theappendix and it’s role in the context of saving a routine ECHO can be seenin Figure 3.1.403.2.2 Data Acquisition and Pre-ProcessingTo perform the data acquisition, the following steps were taken with thehelp of VGH’s Information Technology (IT) team:1. Replicated and anonymized XceleraTM (Philips Healthcare, Nether-lands)’s mySQL instance.2. Installed the new mySQL instance on a computer located at Vancou-ver Coastal Health’s IT department.3. Queried the mySQL database to return all studies that contained seg-mentation information.4. Copied all available ECHO studies from VGH’s Picture Archiving andCommunication System (PACS) system with segmentations that werenot archived.Once all of the data had been acquired, each ultrasound study ECHOinformation was then matched with its corresponding segmentation infor-mation, which was then reconfigured into MATLABTM (Mathworks Inc.MA, USA) structures. Ultrasound images were matched with their correctsegmentation using a unique study identification key. Each saved structurecontained the volume ECHO information, manufacturer name, filename, dateof study, Medical Record Number (MRN), DICOM header information, seg-mentation coordinates and the segmentation frame number for both AP4 andAP2 views. Each ECHO volume was then anonymized by applying a 2D rect-angular black-out region to the US header that was burned into the imagewith confidential information. Furthermore, since LA segmentations wereperformed using the LV measurement module in XceleraTM (Philips Health-care, Netherlands), LA ECHO studies had to be visually separated from stud-ies with LV segmentations to ensure the correct data was used.For this thesis, the data acquired from VGH was filtered by the US ma-chine type. The data acquired in this work was all from Philip’s ultrasoundmachines. Reasons for this include the majority of complete ECHO stud-ies were performed with this manufacturer, imaging quality of this machine41Pixel Spacing (cm)0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.06Number of Studies050010001500200025003000AP2AP4Figure 3.2: Distribution of pixel spacing for each apical view.rendered the most favourable ECHO images of the anatomy, and measure-ments were quantitatively considered as more reliable by VGH sonographerswith this machine. The ECHO data used was also normalized to ensure equalpixel spacing between the images as pixel spacing between two unique stud-ies was variable due to configuration settings and different operators. Themean pixel spacing was 0.364 ± 0.06 mm and 0.386 ± 0.05 mm for theAP4 and AP2 views, respectively. Each AP4 image was resized to have apixel spacing of 0.605 mm and a pixel spacing of 0.658 mm for AP2 images(Figure 3.2).3.3 Patient MeasurementsAll patient measurements are stored on VGH’s FileMakerTM Pro 6 database.Sonographers are required to manually enter all their findings from an ECHO42study along with their patient’s personal information. A brief overview ofthe patient characteristics and general statistics within the database can beseen in Tables 3.1 and 3.2.Table 3.1: Patient ECHO data characteristicsAverage Male Age 66 ± 16 yearsAverage Female Age 65 ± 16 yearsAverage Total Age 66 ± 16 yearsAverage Male BSA 1.96 ± 0.22Average Female BSA 1.69 ± 0.20Average Total BSA 1.83 ± 0.25Average Male LA Volume 39.5 ± 15.03 mLAverage Female LA Volume 37.84 ± 14.92 mLAverage Total LA Volume 38.72 ± 15.00 mLNormal EF dysfunction (65±10%) 71.5%Lower Limits EF dysfunction (50-65%) 13.1%Mild EF dysfunction (50±10%) 4.7%Mild-Moderate EF dysfunction (35-50%) 3.4%Moderate EF dysfunction (35±10%) 2.4%Moderate-Severe EF dysfunction (20-35%) 2.1%Severe EF dysfunction (20%) 0.7%Male Studies 3566Female Studies 3134Unspecified Gender Studies 317Number of Studies 7017Number of Unique Patients 5943Number of Followup Studies 10743.3.1 FileMaker SoftwareFilemakerTM is a relational database program that allows users with min-imal technical knowledge to create a GUI with a backend database. Thissoftware fills a niche role within the cardiology department. Instead ofusing XceleraTM (Philips Healthcare, Netherlands)’s cardiology suite, thecustom FilemakerTM was created to allow advanced searching by not only43Table 3.2: FilemakerTM database data characteristicsTotal Distribution of normal Diastolic Dysfunction (DD) 16%Total Distribution of mild DD 14%Total Distribution of moderate DD 4%Total Distribution of severe DD 1%Total Distribution of indeterminate DD 14%Total Number of Sonographers 120Total Number of Cardiologists 6MRN, but by patient physiology as well. This allows FilemakerTM to be avaluable teaching tool, as physicians can retrieve studies based on certainkeywords.The software operates by allocating four main tabs for each ECHO study.The tabs Front Page, Valves/PA, Aorta/Atria/Shunts Pericardium are theonly pages filled out by the sonographer. There are check boxes and emptyvalues for their report. The last tab, LV/RV/Conclusions is filled out by thecardiologist after reviewing the first three tabs and the ECHO images.3.3.2 Data Acquisition & Data Pre-processingData acquisition from FilemakerTM was performed by first exporting thedatabase into a Comma-Separated Values (CSV) file, where each row repre-sented one study and each column represented a unique field. This methodof export was the only method available at the time of retrieving data for thisthesis. This CSV file was then imported into MATLABTM (Mathworks Inc.MA, USA) using its bulk importer tool, which was required to import over207020 rows and over 450 columns. The imported file was saved as a ma-trix sorted by study date. Once the import process was complete, a for-loopwas used to match a patient’s MRN and date of study in both the ECHO studypreviously gathered and the measurements obtained from the FilemakerTMdatabase. Specific measurements that were matched to each MATLABTM(Mathworks Inc. MA, USA) structure includes LA volume, MR severity,44102030405060708090Combined GendersIndex Volume (mL/m2)20406080100120140Male FemaleVolume (mL)Figure 3.3: (Left) Distribution of left atrial volume for each gender in thedataset. (Right) Distribution of index atrial volume within the dataset.DD level, filling pressure and BSA. The distribution of LA volume from thegathered studies has been plotted in Figure 3.3 and an example image withits corresponding segmentation can be see in Figure 3.4.45Student Version of MATLABFigure 3.4: An example AP4 ECHO from the database with it’s correspondingsonographer’s LA segmentation information. The LA’s centroid has alsobeen identified using the center of weight from the ground truth.46Chapter 4MethodsThe proposed joint information framework for ECHO analysis consists offour phases of view processing, segmentation, and LA volume estimationand disease detection. The objective of view preprocessing is to automati-cally differentiate between AP4 and AP2 views and pinpoint the LA anatomy.In the segmentation phase, we estimate the LA boundary in both AP4 and AP2views. In the volume estimation and disease classification phase, we use thesegmentation result to estimate the LA volume and also, detect and classifyMR severity. In the following, we provide details for each phase.New PatientEcho Information Disease DetectionLA Volume EstimationView Preprocessing LA SegmentationFigure 4.1: Overview of the components of the framework. First, a prepro-cessing step is performed on the acquired echo data to identify and cropthe left atrium region for the following segmentation phase. Next, si-multaneous echo segmentation, volume estimation evaluation and MRdisease labels are estimated from localized apical views.47Training TestingROI LocalizationNew Patient1, …, MApical View ClassificationReconstructionTesting Score VectorsTraining AP4 Score VectorsTraining AP2 Score VectorsPCAAP4AP2Estimated Centroid View PreprocessingFigure 4.2: Overview of the PCA preprocessing phase for the framework.The preprocessing phase has two components of training and testing.The input to the training component is a set of AP4 and AP2 views andtheir associated binary clinical LA segmentation. A set of score vectorsare determined via PCA that reconstruct AP4 and AP2 views and arestored in the memory.4.1 View PreprocesingThe proposed view preprocessing phase, as seen in Fig. 4.2, automaticallyseparates AP4 and AP2 views for the subsequent segmentation phase anddetermines the ROI defining the anatomical location of the LA. In a trainingstep, we use PCA on a data set of AP4 and AP2 views given its computationalefficiency and lowmemory requirements. This allows us to reconstruct eachview from the linear combination of orthogonal basis functions. The scorevectors of such reconstructions along with the LA segmentation contours arethen used in the testing step to perform apical view classification and ROIselection on a new ECHO image. In the subsequent phase for a new queryimage, the following two steps are followed:Apical view classificationThe image is first labeled as either exclusively AP4 or AP2. This is achievedby projecting the query image, using PCA, in the space that is spanned byAP4 and AP2 orthogonal basis functions, and then computing the score vec-tors in each space. Subsequently, the L2 norm between the score vectorsand those stored for all training data, is computed. The image resulting in48AP Imagewvi×ith Coefficient Mixing Element=#ICi = 1ith Principal Component( tv i )Txmmth Patient ObservationA: Mixing Atlaswvi=( tv i )TMinimum Euclidean Distance×ammth Mixing Atlas VectorQuery ImageOfflineOnlineImage Coefficient MatchFigure 4.3: Overview of the apical view classification process. Offline, wecalculate the coefficient scores from PCA on the AP4 and AP2 intensityimage information. Each image’s score is stored in the mixing atlas.Online, we compare each image’s scores stored mixing atlas to the queryimage and calculate the most similar ECHO.the smallest norm determines the associated view and closest match to thequery image.4.1.1 ROI LocalizationTo determine a region of interest containing LA in either the AP4 or AP2view, we use the binary clinical LA segmentation associated with the clos-est match image to estimate the centroid of LA. Subsequently, a fixed-sizebounding box, which was determined heuristically from the training dataand is associated with the largest appearance of LA in the data, is centeredon the estimated LA coordinates. We further refine the estimation of LA lo-cation through an intensity-based registration, where the fixed image is theROI in the query image, and the moving image is the average mean inten-sity of all LA appearances with the same bounding box dimensions in thetraining data. We use an affine transform and mutual information as thesimilarity metric to perform the registration.49ReconstructionTrainingJoint ICA …TestingModel 1 … Model NTesting ICA Concatenated Mixing Coefficients1NPCA2…Model 1 … Model NTraining ICA Concatenated Mixing CoefficientsSegmentation SVM ClassifierReconstruction with Joint ModelsSegmentation SVM ClassifierDisease SVM ClassifierDisease SVM ClassifierLA Segmentation, Volume Estimation and Disease DetectionPatient GroupingJoint AP4 & AP2 Model 1Joint AP4 & AP2 Model 2Joint AP4 & AP2 Model NDisease LabelDisease Label…AP4AP2LA ROIsSegmentations & VolumeFigure 4.4: A schematic diagram of the LA segmentation, volume estimationand disease detection phases. During training, patients are groupedinto clusters based on their LA appearance similarity. Subsequently, ajICA model is generated for each subgroup, and two SVM classifiersare trained. One SVM classifer is used to select the closest model to thequery image and another performs disease classification. During testing,the jICA mixing coefficients from the reconstruction of image intensitymaps in apical views are used to determine LA segmentation contours,compute LA volume, and assign the disease label for an unseen image.4.2 LA SegmentationThe LA ROIs are used within a machine learning framework to segment theLA. In the training step, we cluster the LA contours as labelled by expertsin the clinical data to sets of similar shapes, where a jICA model is con-structed for each cluster (sec. 4.2.2). Subsequently, a segmentation SVMclassifier is trained on the mixing coefficients (see Section 4.2.2) associatedwith each cluster. In the testing step, we project a query image on to thesemodels to perform segmentation and choose the optimal model using SVM.A schematic of the LA segmentation phase is provided in Fig. 4.4.504.2.1 Training: Patient GroupingThe clustering is performed to create detailed models of the main anatomi-cal variations. Initially, we use PCA on the binary LA contours in both AP4and AP2 views to compute score vectors of the training data: 1) We rigidlyalign the contours by their centroid and assume the contours’ correspond-ing orientation is consistent across the dataset based on clinical guidelines;2) We quantize the contour points in polar coordinates. Using the centerof gravity of each contour, we sub-sample the contour at 1 sampling in-terval to convert the contour to a set of points; 3) We perform PCA on thepoint-set across the training data and determine the principal modes of vari-ation [31]. Subsequently, we divide the distribution of scores for the firstprincipal component to N subgroups, determined experimentally.4.2.2 Training: Joint Model GenerationFor each subgroup, we use jICA to train a model that captures the jointspace of LA regions of interest and their corresponding segmentation. Weassume that there is a relation between the intensity variation in ROIs of LAin ECHO, and their binary clinical segmentation. This is not an unreasonableassumption since sonographers use the contrast of the boundary intensityagainst the blood pool to draw the LA contours. The input observations tojICA are the combined intensity information of both apical views (AP4 andAP2), along with their binary clinical segmentation. jICA can be used toidentify any joint set of independent sources that describe the inter-relationbetween intensity and binary contours. Considering the generative modelX = AB (4.1)where X = [IAP4, IAP2,SAP4,SAP2] is the observation matrix of intensity in-formation, IAP4 and IAP2, in an ROI containing LA in each apical view (con-verted to a vector form), and the corresponding segmentation, SAP4 andSAP2. A is a matrix of mixing coefficients (also referred as ICA loadingparameters, or the modulation profile). B is the matrix of joint sources and51=AP4 Image  AP2 Image PatientsX: ObservationsA: × ST: Joint Sources= ×Mixing Matrix#ICi = 1Ai: ith Mixing CoefficientsBiT: ith Joint SourceAP4 Label  AP2 LabelFigure 4.5: Schematic of the jICAmethod. The observation matrix X is madeby concatenating AP4 and AP2 intensity and segmentation pixels in ROIsthat contain the LA in each apical view. jICA maximizes the indepen-dence among the constructed joint sources B, assuming that they sharethe same mixing coefficient matrix A.has the form of B = [B1,B2,B3,B4], where B1 and B2 are the independentsources for intensity information, and B3 and B4 are the independent sourcesfor segmentation information in AP4 and AP2 apical views, respectively. Theaim of jICA is to find the matrix W = A1 so that the estimation U =WXis close to B [18]. We use a MATLAB implementation of jICA availableonline1. A schematic of the jICA approach is shown in Fig. 4.5.The joint sources of B are generated by the logistic InfoMax ICA algo-rithm [7], which is based on a neural network that uses mutual informationminimization to output the number of sources specified.We generate a jICA model for each patient sub-group we have deter-mined in our data using PCA analysis. After training, we do not store theoriginal intensity or segmentation data. The training data knowledge isstored within the compact jICA sources, allowing us to perform segmen-tation quickly and with reduced f tware524.2.3 Training: Segmentation SVM ClassificationThe concatenated jICA mixing coefficient matrices Ai, i= 1, . . . ,N and theirassociated patient subgroups are used to train a segmentation SVM classifier.This classifier maps the concatenated mixing coefficients to subgroup labels1, . . . ,N. Out of the N segmentation contours derived from jICA models, weaim to choose the most accurate segmentation, which by definition is thesegmentation that best matches (based on the highest Quantitative Dice In-dex (DICE) value) between the estimated and gold standard segmentations.Furthermore, the DICE similarity metric is defined below [36]. Where a andb are the estimated and gold standard segmentations:DICE =2|a \b ||a|+ |b | (4.2)4.2.4 Testing: LA Segmentation Based on Each jICA ModelThe process to generate a segmentation from a query image is a fast matrixoperation to solve the system of equations U=WX, to determineW so thatU is the closest approximation of B. Since the query image only containsthe intensity information, the system of equations will only optimize forWby considering the intensity sub-matrices of U and X . Since the mixingcoefficients in A=W1 are the same for the intensity and binary segmenta-tion sources, we can use the computed coefficients from the intensity mapsto efficiently compute an estimated segmentation probability map of the LAcontour in each view. Subsequently, we use a post-processing step to con-vert the estimated map to a final segmentation using a threshold calculatedduring the training phase. The threshold is determined experimentally onthe training folds to maximize the overlap between the estimated LA seg-mentation and their gold-standard clinical segmentation.534.2.5 Testing: Final LA SegmentationTo determine which of the N segmentation to choose from, we use the seg-mentation SVM classification. Based on the estimated concatenated mixingcoefficient matrix A from the query image, the classifier outputs the label ofthe patient subgroup that best represents the LA boundary.4.3 Volume EstimationThe atrium volume estimation is calculated by analyzing the final segmen-tation for both apical views. We calculate the estimated volume by usingthe preferred standard Area-Length 2D method:VLA =83p⇥ (A4C⇥A2C)min(L4C,L2C). (4.3)This technique uses the apical chamber areas (A4C,A2C) and the atriumlong-axis length (L) [53]. Clinically, the LA long-axis dimension is definedas the distance perpendicular from the center of the mitral valve annulus tothe apex of the atrium. However, in practice, we have observed that the ma-jor axis varies slightly from perpendicular to account for unusual anatomy.To account for this variation, we fit an ellipsoid to the LA segmentation ineach apical view, and calculate the length of its major axis. We use thislength to estimate L for both views.4.4 Disease SVM ClassifierThe framework is also used to train a heart disease classifier to learn theassociation between image intensity and its segmentation, with a diseaselabel. We use mitral regurgitation as a proof of concept to distinguish twotypes of mitral regurgitation disease labels: healthy (without MR) and mod-erate/severe MR. We sought to demonstrate this by classifying two typesof patients, those without MR (healthy) and those with moderate or severeMR. We design our classification pipeline to analyze the mixing coefficientsfrom the reconstruction with each jICA model, which produces a concate-54nated matrix of jICA mixing weights. The concatenated matrix of jICA mix-ing weights and disease labels are used to train the disease SVM classifier.During testing when a diseased or healthy patient is introduced to the frame-work, we input the estimated concatenated matrix of jICA mixing weightsinto the trained SVM and estimate their MR disease label. A schematic dia-gram of this process is shown in Fig. Diastolic FunctionIn the previously mentioned methods, from raw ECHO images, we calcu-late volume information and separate healthy from moderately disease MRlevels without any additional information. However, we add an additionalprocess to our framework to supplement clinical knowledge. We incorpo-rate standard measurements performed by a sonographer that are routinelyperformed during a clinical examination protocol to diagnose diastolic dys-function. In a standard ECHO examination there are over 100 common mea-surements to extract correlation from. We reduce this feature set to onlyencompass the relevant features necessary for diastolic dysfunction. Weused the mRMR feature selection algorithm [76], commonly used to deter-mine pairing of genes and phenotypes, to remove redundant features. Weuse mRMR to maximize the joint distribution between the feature set (S) andthe disease label (c); reducing m features, xi, to a compact set of 15 fea-tures, seen in (Table 4.1). Where the mRMR criterion (4.6) is a combinationof maximum relevance (4.4) and minimum redundancy (4.5) of the featureset S and the target class c.max D(S,c),D=1|S| Âxi2SI(xi;c) (4.4)min R(S),R=1|S|2 Âxi,x j2SI(xi;x j) (4.5)max f(D,R),f = DR (4.6)55Where the mRMR criterion (4.6) is a combination of maximum relevance(4.4) and minimum redundancy (4.5) of the feature set S and the targetclass c. (4.4) approximates max D(S,c) using the mean of all mutual in-formation between clinical measurements and the target disease label. Dueto the nature of the maximum relevancey term (4.4), discovered featurescould be highly dependent amongst each other. This is fixed by the min-imum redundancy term (4.5), which removes highly dependent featureswhile minimally affecting the discriminative power of the features. ThemRMR criterion is relevant to apply in echocardiography due to high corre-lation between features and reducing the number of redundant features thatclinicians need to measure. This reduced feature set was then used to traina multi-label SVM classifier from over 10,000 patient records to performclassification of diastolic function and LV filling pressure.56Table 4.1: A subset of common ECHO measurements featuresContinuous ClassificationDiameter of ascending aortaThe fraction of outbound blood pumped from the heart (EF)Posterior wall Thickness at end-diastole (PWd)Passive mitral inflow velocity LV filling (E)Active filing with atrial systole (A)Mitral annular velocities of passive LV filling (E’)The fraction of mitral inflow / mitral annular velocity (E/E’ Ratio)Length of the sinuses within the valsalvaDiameter of the inferior vena cava (IVC)Fraction of the PWd and LV end diastolic diameter (RWT)Right Ventricle diameter in diastole (RVd)Discrete ClassificationMitral regurgitation severityTricuspid regurgitation severityLevel of output of left ventricleAortic regurgitation severityLeft ventricle mass indexRight ventricle function57Chapter 5ExperimentsTo evaluate the presented joint segmentation and labeling framework, weperform a 4-fold cross validation on 6,993 studies that were acquired fromthe routine cardiology care (see Table 3.1) with ethics approval from Clini-cal Medical Research Ethics Board of Vancouver Coastal Health. Through-out the analysis, a rigorous process of data anonymization, de-identificationand data encryption was followed based on the guidelines recommendedby the Vancouver Coastal Health Privacy Office. No special treatment wastaken for handling pathological cases. We used the end-systolic frame ofeach acquired ECHO cine, its clinical LA segmentation contour for AP4 andAP2 views, and MR disease labels in the archival data as the gold standard.To normalize the data, images were scaled to the largest pixel resolutionin the dataset. The choice of parameters for our framework is detailed inSec. 5.1. The evaluation was performed with respect to the apical viewclassification (Sec. 5.2), the LA volume estimation accuracy (Sec. 5.4), theMR classification accuracy (Sec. 5.5), diastolic dysfunction and LV fillingpressure classification (Sec. 5.6), and the computational complexity (Sec.5.7).585.1 Framework ParametersFor each phase of the framework, we have set global parameters for thementioned methods above. In the offline phase, we stratify patients intoone of seven groups (determined experimentally), based on the histogrambin their mixing coefficients belong in. The joint model was optimized us-ing 7 patient subgroups. These seven sets of patient studies are used to trainseven joint segmentation models using jICA. When using jICA, unlike PCA,choosing the number of components that are needed to best describe an ob-servation matrix is not well defined and can affect results [61]. We followthe idea used in functional magnetic resonance imaging cognitive studiesto estimate the number of optimal sources. There has been success usingMinimum Descriptive Length (MDL) criterion [57] to estimate the numberof sources. We use the local minimum from MDL information-theoretic cri-terion to determine the number of independent sources to represent the eachmodel (Fig. 6.6). During SVM training, we use the The LIBSVM classifiertoolbox to perform classification of the optimum segmentation model andMR labels. The SVM classifier’s hyper-parameters are tuned using all threetraining folds using a Gaussian Radial Basis Function (RBF) kernel. Thisresulted in the following parameters for model selection: C= 4 (regulariza-tion parameter), g = 0.125. For MR selection, we train the classifier, usingC= 2, g = 0.0625, by grouping healthy and mild MR patients together in oneset and patients with moderate and severe MR in another. We determinedthese parameters via an exhaustive search to maximize classification accu-racy. During the online evaluation, we use a 4-fold cross validation (threefolds for training, one fold for testing). In the localization process, the in-tensity based registration has a minimum step length of 0.025, maximumstep length of 0.0425 and a limit of 300 iterations.5.2 Apical View ClassificationWe perform an experiment on the validation dataset to determine the ac-curacy of view classification, using the known AP4 and AP2 labels as the59gold-standard. We report the classification accuracy available when detect-ing each view.5.3 ROI LocalizationThe LA localization error was determined as the Euclidean distance betweenthe calculated centroid location of the LA bounding box and its gold stan-dard position in the validation data.5.4 Segmentation and Volume Estimation of the LAThe accuracy of the LA volume estimation was determined as the Pearsoncorrelation coefficient and the Sum of Squared Errors (SSE) compared tothe clinical gold standard measurement. Note that a direct comparison ofmaximal atrium volume with 3D volume estimation was not possible, as ourdataset had no 3D LA measurements. Furthermore, we investigated the in-fluence of using a joint model (AP4+AP2) incorporating both apical views,compared to using only a single model for each apical view, on the segmen-tation accuracy.5.5 MR Diagnostic LabelsIn a cohort of 1007 patients, including 424 healthy patients and 583 moder-ate or more severe MR we generate the jICA model on the combined trainingsets obtained using a 4-fold cross validation to ensure stability of the gener-ated joint components.5.6 Diastolic Dysfunction and LV Filling PressureFurthermore, the framework’s component to predict diastolic dysfunction isperformed by analyzing our large data access of 10490 patient measurementreports to provide database-guided recommendations. In our local hospital,the patient’s diastolic function grade distribution was found to be: normal(16%), mild (14%), moderate (4%), severe (1%) and indeterminate (14%).60To predict diastolic dysfunction we divided the reports into training (80%)and testing (20%) and trained a multi-set SVM classifier based on the fea-tures shown in Table 4.1. We use five-fold validation to test the stability ofthis trained SVM.5.7 Time and Computational ComplexityRuntime was measured in seconds for four independent groups of computa-tions run on a standard PC (Intel Core i7, 2.93GHz, 8GB RAM): 1) modelsand SVM classifier training; 2) left atrium localization; 3) estimation of theleft atrial volume from segmentation; and 4) SVM classifier application foroptimal model selection and disease labels. The framework assumes thatthere are two orthogonal ECHO inputs of the heart that are in adherence withthe clinical standard AP4 and AP2 orientations. The computational complex-ities of the testing phase to be considered are using an affine intensity-basedregistration, two SVM classifiers, and PCA/jICA reconstruction.61Chapter 6Results and DiscussionIn this thesis, we have introduced a joint information framework for fastand computationally efficient ECHO analysis. The core of our method is theapplication of jICA on a combined data matrix of echo images and their cor-responding segmentation labels for simultaneous LA segmentation in twoapical ECHO views and MR disease detection. The framework is generic andcan potentially be extended from 2DE to 3DE, as long as a large retrospectivepatient database is available. The flexibility and true power of the frame-work is seen by combining intensity and segmentation information into aunified space to reveal diagnosis labels from only ECHO image information.By using a multi-set jICA approach, we have introduced a framework ca-pable of learning from the observable correlation between each ultrasoundintensity pixel and the corresponding segmentation labels.Additionally, outside of this framework, we have investigated featureselection to improve disease classification for diastolic dysfunction and LVfilling pressure. By introducing two classifiers, which perform with goodaccuracy, we show potential to provide labels to prevent patient parametersfrom being labelled as indeterminate.62Distance between Estimated LA centroid and the Gold Standard (mm)0 1 2 3 4 5 6 7 8Number of Studies01000200030004000500060007000(a) Initial AP4 ROI localization errorDistance between Estimated LA Ceentroid and the Gold Standard (mm)0 1 2 3 4 5 6 7Number of Studies0100020003000400050006000(b) Initial AP2 ROI localization errorFigure 6.1: Initialization error between the estimated LA centroid and thegold standard. Determined by finding the most similar ECHO image.6.1 Apical View Classification and LocalizationApplying our joint information framework to the analysis of n=6993 ECHOdatasets, we achieved a view classification accuracy of 87% and 99%, forthe AP4 and AP2 views, respectively. Compared to AP2 classification rate,the AP4 classification rate was likely lower due to the variable noise in theimage, which could have reduced the visibility of the right ventricle andatrium, causing the AP4 to resemble the AP2 anatomical structure. In ourframework, we assume that there is only one unseen AP4 and one unseen AP2image, which allows us to accurately identify the AP2 image and classify theremaining image as the AP4.Furthermore, we simultaneously determine the ROI defining the anatom-ical location of the LA. Our initial localization accuracy of the LA in the AP4was 3.75 ± 3.1 mm and 4.37 ± 3.3 mm in the AP2 view (Figure 6.1). Thisaccuracy was improved using a refinement affine registration step, achiev-ing an ROI localization accuracy was 3.2 ± 3.0 mm in AP4 and 3.5 ± 3.3mm in AP2 ECHOs (Figure 6.2).63Distance between Estimated Centroid and the Gold Standard (mm) 0 1 2 3 4 5 6 7Number of Studies01000200030004000500060007000(a) Refined AP4 ROI localization errorDistance between Estimated LA Centroid and the Gold Standard (mm)0 1 2 3 4 5 6 7Number of Studies01000200030004000500060007000(b) Refined AP2 ROI localization errorFigure 6.2: Refined centroid estimation results for AP4 and AP2 using anaffine intensity-based registration to an average LA appearance.6.2 Segmentation and Volume Estimation of the LAThe computed volume of the LA resulted in a volume Pearson correlationcoefficient of R = 0.87, a SSE of 12 mL (Figure 6.3), and an average DICEcoefficient of 0.91 and 0.90 (AP4 and AP2, respectively) when compared tothe clinical gold standard. On the other hand, using the single segmentationmodel yielded AP4 and AP2 DICE coefficients of 0.87 and 0.86, respectively(Figure 6.4). Figure 6.5 shows 10 example segmentation cases of both theAP4 and AP2. The joint model was optimized using 52 joint sources, unlikethe single models, which both used 75 joint sources (Figure 6.6).To the best of our knowledge, our validation dataset containing approxi-mately 7,000 patients is the largest of its kind to date. While there are manysegmentation methods specific to the left ventricle, segmentation of LA fromECHO data has been only recently reported in a small cohort of 10 patientson AP4 views [19]. We suspect that this is due to the high variability and thepoor signal-to-noise ratio of LA in ECHO, as it is the farthest chamber awayfrom the ECHO sweep. A recent report also confirms the challenges associ-ated with LA segmentation in 3D Magnetic Resonance Imaging (MRI) andCT data [88]. On a limited dataset of 30 CT and 30 MRI images (divided into640 50 100 150 200 250 300050100150200250300SSE=12 mLr=0.8683Gold Standard Volume (mL)Estimated Volume (mL)0 50 100 150 200 250 300−150−100−50050100150Mean Gold Standard Volume & Estimated Volume (mL)Estimated Volume − Gold Standard Volume (mL)17 (+1.96SD)−11 [p=0]−39 (−1.96SD)RPC: 28 mL (43%)CV: 22%  TitleGNMAESStudent Version of MATLAB0 50 100 150 200 250 300050100150200250300SSE=12 mLr=0.8683Gold Standard Volume (mL)Estimated Volume (mL)0 50 100 150 200 250 300−150−100−50050100150Mean Gold Standard Volume & Estimated Volume (mL)Estimated Volume − Gold Standard Volume (mL)17 (+1.96SD)−11 [p=0]−39 (−1.96SD)RPC: 28 mL (43%)CV: 22%  TitleGNMAESStudent Version of MATLABFigure 6.3: (Top and bottom): The correlations between estimated and goldstandard LA volumes and the corresponding agreement between themanual and estimated methods are shown in the Bland-Altman plot.650.50.550.60.650.70.750.80.850.90.951Single Model AP4 Joint Model AP4 Single Model AP2 Joint Model AP2DICEStudent Version of MATLABFigure 6.4: Comparison of segmentation accuracies (measured by DICE co-efficient) for segmentations only based on models constructed from aspecific view (single model) and on models jointly derived from multi-ple apical (joint) views.10 training and 20 validation data cases), DICE coefficients were compara-ble to those we report in this work. The literature is unclear as to how accu-rate left atrial estimations are required to be, especially when compared to3DE technology [44]. Furthermore, the European Society of Hypertensionrecommends that LA volumes should be thresholded at 34 ml/m2 (a valuethat implies diastolic dysfunction) when defining an abnormal LA size forECHO, MRI and CT.The model selection in LA segmentation has a direct impact on the vol-ume procured. Visual assessment of the first principal compo ent used forpatient clustering shows that this component is directly related to the LAsize. A mismatched model has the potential to over- or under-estimate vol-umes as shown by the outliers in Figure 6.3. The segmentation SVM modelselection failed in 8% of patient cases where the length of the major axis,66(a) (b)(c) (d)(e) (f)(g) (h)(i) (j)Figure 6.5: Example studies of estimated segmentation results (dashed).670 200 400 600 800 1000 120011.522.533.544.5 x 106  Single ModelCombined ModelStudent Version of MATLABNumber of Independent Components Log Likelihood Criterion(Arbitrary Units)Figure 6.6: MDL Criterion results demonstrating the optimal number of jointsources for the single and joint segmentation models.derived from the segmentation of the AP4 and AP2 views, were not within20% of each other, in accordance to clinical practice [48]. Within the failedcases set, 25% had been marked by physicians as having sub-optimal im-age quality, 15% had a gold standard major axis deviation of over 20%,and 3.5% had mitral valves prosthetics installed. We expect that the modelselection could be improved by incorporating the assessment of additionalstandard ECHO views of the LA into the framework. To visualize the infor-mation from mixing coefficients used in the segmentation SVM classifier,we use tSNE [34] to visualize one fold of the high-dimensional AP4 datasetin Figure 6.8. Using the Barnes-Hut implementation of tSNE1, we visu-alize of the dataset mixing coefficients through it dimensionality reductiontechnique.We can see the separation of the larger blood pools (bottom) com-pared to those with smaller LA size (top). We can also observe the high vari-ability of the LA’s shape from circular to an oblong ellipsoid in Figure 6.8,demonstrating the difficulty of the model selection process. Furthermore,tSNE has grouped LA’s with shadowing due to mechanical valves togetherin the bottom right. The shape and appearance of special cases like theseare hard to characterize and can mask the disease’s perceived severity.1htt ps : // Example AP4 Segmentation Shapes(b) Example AP2 Segmentation ShapesFigure 6.7: LA segmentations grouped in columns based on their first prin-cipal components. Clustering is done to improve accuracy, given thediversity of our large dataset.69Figure 6.8: t-SNE is used to visualize the separability of AP4 joint ICA recon-struction weights of a random subgroup of 400 patients. This approachmodels the high-dimensional space in a two-dimensional system, suchthat nearby images are have similar AP4 anatomy and dissimilar AP4 im-ages are further away. Images highlighted in red are those with severelyenlarged left atriums (40 mL/m2)70Figure 6.9: An example of LA shadowing in the ultrasound created by a me-chanical mitral prosthesis valve.6.3 MR Diagnostic LabelsWe use our framework to also detect patients with moderate or severe MRwhich is among the most frequent valve diseases resulting in LA enlarge-ment and ultimately in LV dysfunction. For classification of MR diseaselabel, we achieved a classification accuracy of 82.3%, specificity of 70.6%and a sensitivity of 91.6% (Figure 6.10). Without proper treatment, symp-tomatic patients have an annual death rate of over 5% [20]. MR is mainlycharacterized by the reverse blood flow from the LV to the LA and is thusmost commonly assessed with respect to its mechanism and severity usingnoninvasive color Doppler [108]. However, Doppler ECHO measurementsare based on geometric assumptions of the regurgitant jet, increasing thepotential for misclassification. Current observer variability in detecting MRusing Doppler is reported to account for approximately 25% [4]. By usingLA dynamics to detect MR before it becomes unmanageable, we introduceda useful tool that allows physicians to alleviate diastolic dysfunction andsafely manage heart failure. The ability to detect if a patient suffers frommoderate MR during a routine examination only using B-Mode ECHO imag-7149% 8% 15% 44% Predicted Diagnosis  True Diagnosis  p n p’ n’ 86.11 ± 0.04% 74.10 ± 0.05%   79.90 ± 0.04%  Sensitivity Specificity Accuracy 82.24 ± 0.06% 69.67 ± 0.06%   75.85 ± 0.03%  With Computed Volume Without Computed Volume 49% 8% 15% 44% Predicted Diagnosis  True Diagnosis  p n p’ n’ 86.11 ± 0.04% 74.10 ± 0.05%   79.90 ± 0.04%  Sensitivi y Specificity Accuracy 82.24 ± 0.06% 69.67 ± 0.06%   75.85 ± 0.03%  With Computed Volume Without Computed Volume 91.6%0.1%82.3%53449129295Figure 6.10: The confusion matrix for the healthy-moderate/severe and cor-responding is of high clinical importance because it provides a means to triggerfurther examination visits to follow up the development of the disease.Using our framework, we can classify patients with moderate to severeMR disease purely based on B-Mode ECHO image information. To under-stand the inputs used to train the disease SVM classifier, we visualize thelearned independent sources in Figure 6.11. Based on the highlighted areasof modulation in Figure 6.11, we conclude that the sources grow and shrinkbased on an unseen TTE’s ultrasound intensity information, thus capturingthe size of the atrium. As a result, our disease SVM classifier is likely learn-ing to discriminate between normal and enlarged left atriums for MR detec-tion. Clinically, using this information as a classifier is appropriate since,regardless of Doppler grading, severe chronic MR does not exist withoutleft atrial enlargement [79]. Additionally, we compare the abilities of thelearned sources for the joint and single model for MR classification, how-ever, the mixing weights reconstructed from either jICA modelling approachhad no apparent advantage (Figure 6.12). Furthermore, we investigate thecases where we have misclassified patient labels. We compare patients whowere labeled as False Positive (FP) and False Negative (FN) against the TruePositive (TP) and True Negative (TN) features. First, within the FP and TNsets, we see that there is potential to miss-classify larger patients with in-creased BSA, seen in Figure 6.14(A). This indicates that patients with an72AP4 USAP4 CTSAP2 USAP2 CTSFigure 6.11: Learned joint independent sources. Each source represents aregion-specific modulation in the AP4 and AP2 image intensity and theassociated clinical LA segmentations. Highlighted areas show areaswith more than 60% modulation of the sources. These 15 joint sourceswere created from a joint ICA decomposition of AP4 and AP2 apicalviews from one of the patient sub-groups.overall enlarged anatomy should be more carefully screened after our clas-sification. Secondly, we also compare the LA volume of patients within theTP and FN sets, and note the misclassified patients have a significantly de-creased mean LA volume of 35 ml/m2 (Figure 6.13(B)). This finding is ofclinical interest, as patients with increased levels of chronic MR will have LAdilation (36ml/m2 [95]). Furthermore by investigating the surface area ofthe reguritant valve (Figure 6.15), we observe the FN group has decreasedindexed annular area, suggesting these patients within the FN set could suf-fer acute MR. As chronic MR progresses over time, it enlarges the heart’sanatomy and the diameter of the mitral valve. Acute MR is not normallyreflected in the LV and LA chamber, and can instead manifest with physicalsymptoms.730 0.2 0.4 0.6 0.8 Positive RateTrue Positive RateMitral Regurgitation Classification ROC Curve  Combined Model (AUC = 0.83)Single Model (AUC = 0.82) Figure 6.12: The ROC curve for the healthy-moderate/severe SVM classifierusing a single and joint model. Mitral regurgitation is trained only onreconstructed weights of ECHO features from the estimation process.6.4 Diastolic Dysfunction and LV Filling PressureWe achieved a diastolic dysfunction classification accuracy of 95.1%, speci-ficity of 95.6% and a sensitivity of 80.8% (Figure 6.16). For LV filling pres-sure classification accuracy, specificity and sensitivity were 96.8%, 96.4%and 97.1%, respectively. These results were achieved by using a subsetof features chosen by the Minimum Redundancy Maximum Relevancy se-lection algorithm. With a preset number of cardiac feature parameters, thealgorithm determined the parameters that were mutually far away from eachother while still having a strong correlation to the classification label. Out ofthe most common 46 features determined by experts (found in Figure 4.1),74468101214True Positive False NegativeMR Confusion Matrix LabelMitral Annulus Surface Area (cm2/m2 )Mitral Annulus Surface Area00. Negative True PositiveLA VolumeStudent Version of MATLAB00. Positive True NegativeLA VolumeStudent Version of MATLAB00. Negative True PositiveBody Surface AreaStudent Version of MATLAB00. Positive True NegativeBody Surface AreaStudent Version of MATLABBody Sur ce Area LA lumeBody Surface AreaFalse Negative          True PositiveFalse Negative  True PositiveLA olume False Positive    True NegativeFalse Positive      True Negative(a)468101214True ositiv f i  t iMitral Annulus Surface Area (cm2/m2 )Mitral Annulus Surface Area00. Negative True Pos tiveLA VolumeStudent Version of MATLAB00. Positive True NegativeLA VolumeStudent Version of MATLAB00. Negative True PositiveBody Surface AreaStudent Version of MATLAB00. Positive True NegativeBody Surface AreaStudent Version of MATLABBody Sur ce Area LA lumeBody urface AreaFalse Negative  True PositiveFalse Negative  True PositiveLA olume False Positive    True NegativeFalse Positive      True Negative(b)Figure 6.13: Visualization of influence regarding normalized body surfacearea (a) and LA volume (b) on False Positives and True Negatives forMR classification.468101214True Positive False NegativeMR Confusion Matrix LabelMitral Annulus Surface Area (cm2/m2 )Mitral Annulus Surface Area00. Negative True PositiveLA VolumeStudent Version of MATLAB00. Positive True NegativeLA VolumeStudent Version of MATLAB00. Negative True PositiveBody Surface AreaStudent Version of MATLAB00. Positive True NegativeBody Surface AreaStudent Version of MATLABBody Sur ce Area LA lumeBody Surface AreaFalse Negative          True PositiveFalse Negative  True PositiveLA olume False Positive    True NegativeFalse Positive      True Negative(a)468101214True Positive False NegativeMR Confusion Matrix LabelMitral Annulus Surface Area (cm2/m2 )Mitral A n lus Surface Area00. Negative True PositiveLA Vol meStudent Version of MATLAB00. Positive True NegativeLA VolumeStudent Version of MATLAB00. Negative True PositiveBody Surfac AreaStudent Version of MATLAB00. Positive True NegativeBody Surface AreaStudent Version of MATLABBody r ce Area L  lumeBody rface AreaFalse Negative       True PositiveFalse Negative  True PositiveL  olume False Positive    True NegativeFalse Positive     True Negative(b)Figure 6.14: The affect of normalized body surface area (a) and LA volume(b) patient features on MR classification’s False Positives and TrueNegatives.75468101214True Positive False NegativeMR Confusion Matrix LabelMitral Annulus Surface Area (cm2/m2 )Mitral Annulus Surface Area2468101214Healthy MR Moderate − Severe MRMitral Annulus Surface AreaMitral Surface Area (cm2/m2 )Student Version of MATLABMitral Annulus Surface Area Mitral Annul s Surface AreaHealthy MR Moderate – Severe MRTrue Positive  False Negative(a)468101214True Positive False NegativeMR Confusion Matrix L belMitral Annulus Surface Area (cm2/m2 )Mitral Annulus Surface Ar a2468101214Healthy MR Moderate − Severe MRMitral Annulu Surface Ar aMitral Surface Area (cm2/m2 )Student Version of MATLABitral Annulus Surface Area itral Annul s Surface AreaHealthy R oderate – Severe RTrue Positive  False Negative(b)Figure 6.15: A comparison of the surface area of the mitral annulus betweenpatients with no MR and with moderate and above was experimentally determined that the optimum number of parameterswas 16 and 36, for diastolic dysfunction (Figures 6.1 and 6.17) and LV fill-ing pressure prediction respectively. Indicating that there were numerousparameters that were redundant, such as AP4 and AP2 surface area calcula-tions.The need for diastolic dysfunction classification is not only importantbecause it is one of the criteria for diagnosing heart failure [75], but it’sprevalence in the elderly exceeds 40% [26]. Using the American Societyof Echocardiography (ASE) 2009 guidelines, their are a set of specific pa-rameters that must be met before further grading of the dysfunction cancontinue. In a 2010 study [26], it was found that 47% of 1369 ECHO studiescould not be classified as they did not meet all of the ASE standards. Thus,being able to determine the grading of diastolic dysfunction, and contribut-ing markers like LV filling pressure, is not an objective classification forelderly patients referred for ECHO. By incorporating a more fluid classifica-tion system, based on prior information, there is potential to better definethe dysfunction severity of this group and prevent a high percentage of thepopulation from being classified as indeterminate.76Visual	Confusion	MatrixSevereModerateMildNormalSevereModer-ateMildNormalFigure 6.16: A diastolic dysfunction confusion matrix. Diastolic dysfunc-tion is trained using 16 of the top clinical measurements routinely per-formed. We show a classification distribution heat map confusion ma-trix of diastolic dysfunction.6.5 Time and Computational ComplexityPerforming volume estimation, bi-plane segmentation and classification ona per-patient basis can be completed within 20 seconds, with the registrationbeing the slowest step (approximately 13 seconds) (Figure 6.2. There ispotential to decrease the registration time by running this step in a parallelprocessing code. However, at the time these experiments were conducted,the code was not optimized for speed performance.77Number of Features0 5 10 15 20 25 30 35 40 45 50Classification Accuracy (%)657075808590(a) Diastolic dysfunction feature selection using mRMRNumber of Features0 5 10 15 20 25 30 35 40 45 50Classification Accuracy (%)7880828486889092949698(b) Filling pressure feature selection using mRMRFigure 6.17: Feature selection using mRMR comparing classification accu-racy.78Table 6.1: Diastolic dysfunction mRMR ECHO measurements featuresLateral Doppler measurement of the Mitral Annulus (Lat E’)Septal Doppler measurement of the Mitral Annulus (Sep E’)The mitral inflow Doppler E velocity/ E tissue Doppler velocityLevel of output of Left VentricleDiameter of Ascending AortaMitral regurgitation severityPassive mitral inflow velocity LV filling (E)Posterior wall Thickness at end-diastole (PWd)Tricuspid regurgitation severityFraction of the PWd and LV end diastolic diameter (RWT)Right Ventricle diameter in diastole (RVd)Left Ventricle mass indexAortic regurgitation severityActive filing with atrial systole (A)Length of the sinuses within the valsalvaMitral regurgitation severityTable 6.2: Computational Time (Online)Process Average Time (s)Centroid Estimation and LA Localization 2.4Affine Image Registration 12.44jICA Reconstruction and Post Processing 2.94Segmentation SVM Classifier 0.07Disease SVM Classifier 0.0179Chapter 7ConclusionIn this thesis, we proposed an extensible joint information framework forautomatic estimation of cardiac parameters from TTE data by simultane-ously segmenting AP4 and AP2 ECHO views, which are both required forvolume calculations. We specifically focused on a large cohort of LA ECHOdata, and leveraged the data’s diversity to generate a set of comprehensivemodels using ICA [18] that are capable of assisting in diagnosis. Our frame-work fused ECHO image intensity information and their segmentations frommultiple 2DE views of the heart to automatically estimate clinical parame-ters and diagnostic labels. We segmented both AP4 and AP2 views simulta-neously by using models within our framework that incorporated intensityinformation from both views. Using maximally independent basis functionsto learn the observable patterns between intensity and the correspondingsegmentation, we employed these patterns for LA volume estimation and,as a corollary objective, classification of individuals with chronic MR. Theapproach consisted of the following steps: 1) During training, using a jointindependent component analysis of image intensity information, and thecorresponding labels, associated with clinical measurements and diagnosis,to generate models that jointly describe the image and label space of thosepatients; 2) During evaluation, it segmented the anatomy of interest, andestimated the volume from simultaneous analysis of multiple anatomical80views in echo. 3) For diagnosis, it exploited the generated intensity-labeljoint patterns determined for each abnormal pathology group to classify newstudies. Furthermore, we looked to reduce the number of indeterminate pa-tients who are investigated for diastolic dysfunction and filling pressure;towards enabling standardized patient care.In this work, with the investigated joint information framework focusedon LA estimation and disease label analysis, we made the following contri-butions:• Obtaining a largest of its kind dataset of ECHO information and patientparameters by interfacing and anonymizing VGH’s medical ECHO stor-age system and clinical database.• Proposing a segmentation technique that jointly fuses information frommultiple apical views to aid in overcoming the obstacles of segment-ing the LA region.• Investigating a new technique for classifying normal and moderate ormore severe mitral regurgitation based on jICA reconstruction coeffi-cients. Demonstrating the unique information provided by the jICAlearned sources.• Investigating the use of combining clinical parameters for analysis di-astolic dysfunction and filling pressure to prevent indeterminate clas-sification of patient measurements. Simulations showed the methodcould has potential to be used in parallel with traditional dysfunctionand filling pressure classification.7.1 Future WorkThere is room for future work to improve the accuracy of this framework,namely, improving the patient grouping, enhancing the automatic localiza-tion process, and improving the segmentation SVM classification accuracy.Currently, patients are grouped based on their first-mode of variation from81PCA; that generally corresponds to the overall surface area. In the future,more complex models could be formed from our current database to in-corporate age, body surface area and past medical history. This could beespecially useful in cases where the LA cavity is severely deformed due toold age. Similarly, there is room to improve the localization process, assome ROIs did not contain the entire LA blood pool due to the intensity-based registration method. When performing volume estimation using ROIscontaining LA, aligned based on the gold-standard segmentation of LA inthe query image instead of using the intensity-based registration, we findour method achieves an improved Pearson correlation coefficient of 0.90and a reduced SSE of 11 mL. This result suggests that developing a moreaccurate localization method could improve the framework.Finally, the segmentation SVM model selection failed in 8% of patientcases where the length of the major axis, derived from the segmentation ofthe AP4 and AP2 views, were not within 20% of each other, in accordance toclinical practice [48]. Within the failed cases set, 25% had been marked byphysicians as having sub-optimal image quality, 15% had a gold standardmajor axis deviation of over 20%, and 3.5% had mitral valves prostheticsinstalled. We expect that the model selection could be improved by incor-porating the assessment of additional standard ECHO views of the LA intothe framework.Future work to improve MR classification could include incorporatingDoppler information and the analysis of the entire ECHO sequence insteadof only end-systolic views that we analyze in this work. The effectiveness ofthis approach has been recently demonstrated in 3DE data in a small patientstudy [96].To conclude, while there are many left ventricle segmentation approachesavailable for use in echocardiography, we believe we are the first to proposea joint multi-view segmentation of LA and MR label analysis using B-modeintensity information. The flexibility and true power of the framework isseen by combining intensity information into a unified space to reveal di-agnosis labels from only 2DE image information. The benefits advanced by82this unified framework are: 1) it can estimate the LA volume by learningthe large anatomical diversity of patients witnessed in approximately 7,000studies, or 14,000 images, and obtain results that have high correlation withclinical measurements; 2) It can create patient-specific models capable ofidentifying MR labels; and 3) Due to the low computational complexity ofthe framework, there is potential for our framework to be applied to portableECHO machines that are frequently used in the emergency room and in ruralclinics.83Bibliography[1] W. P. Abhayaratna, J. B. Seward, C. P. Appleton, P. S. Douglas, J. K.Oh, A. J. Tajik, and T. S. M. Tsang. Left atrial size: physiologicdeterminants and clinical applications. Journal of the AmericanCollege of Cardiology, 47(12):2357–63, 2006. ! pages 2[2] M. Alessandrini, T. Dietenbeck, O. Basset, D. Friboulet, andO. Bernard. Using a geometric formulation of annular-like shapepriors for constraining variational level-sets. Pattern RecognitionLetters, 32(9):1240–1249, 2011. ! pages 31[3] E. D. Angelini, S. Homma, G. Pearson, J. W. Holmes, and A. F.Laine. Segmentation of real-time three-dimensional ultrasound forquantification of ventricular function: a clinical study on right andleft ventricles. Ultrasound in Medicine & Biology, 31(9):1143–58,2005. ! pages 36[4] E. E. Apostolakis and N. G. Baikoussis. Methods of estimation ofmitral valve regurgitation for the cardiac surgeon. Journal ofCardiothoracic Surgery, 4(1):34, 2009. ! pages 71[5] E. Aune, M. Baekkevar, J. Roislien, O. Rodevand, and J. E.Otterstad. Normal reference ranges for left and right atrial volumeindexes and ejection fractions obtained with real-timethree-dimensional echocardiography. European Journal ofEchocardiography, 10(6):738–44, 2009. ! pages 3[6] M. E. Barnes, Y. Miyasaka, J. B. Seward, B. J. Gersh, A. G.Rosales, K. R. Bailey, G. W. Petty, D. O. Wiebers, and T. S. M.Tsang. Left atrial volume in the prediction of first ischemic stroke inan elderly cohort without atrial fibrillation. Mayo ClinicProceedings, 79(8):1008–14, 2004. ! pages 384[7] A. J. Bell and T. J. Sejnowski. An Information-MaximizationApproach to Blind Separation and Blind Deconvolution. NeuralComputation, 7(6):1129–1159, 1995. ! pages 52[8] O. Bernard, J. D’hooge, and D. Freboulet. Statistical Modeling ofthe Radio-Frequency Signal in Echocardiographic Images Based onGeneralized Gaussian Distribution. In 3rd IEEE InternationalSymposium on Biomedical Imaging: Macro to Nano, 2006., pages153–156. IEEE. ! pages 34[9] O. Bernard, J. D’hooge, and D. Friboulet. Statistics of theradio-frequency signal based on K distribution with application toechocardiography. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control, 53(9):1689–1694, 2006. !pages 34[10] O. Bernard, B. Touil, A. Gelas, R. Prost, and D. Friboulet. ARBF-Based Multiphase Level Set Method for Segmentation inEchocardiography using the Statistics of the Radiofrequency Signal.In IEEE International Conference on Image Processing, volume 3,pages III – 157–III – 160. IEEE, 2007. ! pages 4, 34[11] O. Bernard, B. Touil, A. Gelas, R. Prost, and D. Friboulet.Segmentation of myocardial regions in echocardiography using thestatistics of the radio-frequency signal. Functional Imaging andModeling of the Heart, pages 433–442, 2007. ! pages 4, 34[12] D. Beymer, T. Syeda-Mahmood, A. Amir, and S. Adelman.Automatic estimation of left ventricular dysfunction fromechocardiogram videos. In Computer Vision and PatternRecognition Workshops, pages 164–171. IEEE, 2009. ! pages 13[13] T. Binder, M. Su¨ssner, D. Moertl, T. Strohmer, H. Baumgartner,G. Maurer, and G. Porenta. Artificial neural networks and spatialtemporal contour linking for automated endocardial contourdetection on echocardiograms: a novel approach to determine leftventricular contractile function. Ultrasound in Medicine & Biology,25(7):1069–1076, 1999. ! pages 4[14] D. S. Blondheim, R. Beeri, M. S. Feinberg, M. Vaturi, S. Shimoni,W. Fehske, A. Sagie, D. Rosenmann, P. Lysyansky, L. Deutsch,85M. Leitman, R. Kuperstein, I. Hay, D. Gilon, Z. Friedman,Y. Agmon, Y. Tsadok, and N. Liel-Cohen. Reliability of visualassessment of global and segmental left ventricular function: amulticenter study by the Israeli Echocardiography Research Group.Journal of the American Society of Echocardiography : officialpublication of the American Society of Echocardiography, 23(3):258–64, 2010. ! pages 2[15] J. G. Bosch, S. C. Mitchell, B. P. F. Lelieveldt, F. Nijland, O. Kamp,M. Sonka, and J. H. C. Reiber. Automatic segmentation ofechocardiographic sequences by active appearance motion models.IEEE Tansactions on Medical Imaging, 21(11):1374–83, 2002. !pages 4, 33, 36[16] T. M. Bui, A. Coron, J. Mamou, E. Saegusa-Beecroft,T. Yamaguchi, E. Yanagihara, J. Machi, S. L. Bridal, and E. J.Feleppa. Modeling the envelope statistics of three-dimensionalhigh-frequency ultrasound echo signals from dissected humanlymph nodes. Japanese Journal of Applied Physics, 53(7 Suppl):07KF22, 2014. ! pages 34[17] T. M. Bui, A. Coron, J. Mamou, E. Saegusa-Beecroft, J. Machi,A. Dizeux, S. L. Bridal, and E. J. Feleppa. Level-set segmentationof 2D and 3D ultrasound data using local gamma distribution fittingenergy. In 2015 IEEE 12th International Symposium on BiomedicalImaging (ISBI), pages 1110–1113. IEEE, 2015. ! pages 34[18] V. D. Calhoun, T. Adali, N. R. Giuliani, J. J. Pekar, K. A. Kiehl, andG. D. Pearlson. Method for multimodal analysis of independentsource differences in schizophrenia: combining gray matterstructural and auditory oddball functional data. Human BrainMapping, 27(1):47–62, 2006. ! pages 5, 52, 80[19] Y. Cao, P. McNeillie, and T. Syeda-Mahmood. Segmentation ofAnatomical Structures in Four-Chamber View EchocardiogramImages. In 22nd International Conference on Pattern Recognition,pages 568–573. IEEE, 2014. ! pages 4, 28, 36, 64[20] B. A. Carabello. Mitral valve repair in the treatment of mitralregurgitation. Current Treatment Options in CardiovascularMedicine, 11(6):419–425, 2009. ! pages 3, 7186[21] G. Carneiro, J. Nascimento, and A. Freitas. Robust left ventriclesegmentation from ultrasound data using deep neural networks andefficient search methods. In IEEE International Symposium onBiomedical Imaging: From Nano to Macro, pages 1085–1088.IEEE, 2010. ! pages 4[22] G. Carneiro, J. C. Nascimento, and A. Freitas. The segmentation ofthe left ventricle of the heart from ultrasound data using deeplearning architectures and derivative-based search methods. IEEETransactions on Image Processing, 21(3):968–82, 2012. ! pages36[23] V. Caselles, R. Kimmel, and G. Sapiro. Geodesic active contours.International journal of computer vision, 22(1):61–79, 1997. !pages 29[24] Y. Chen, H. D. Tagare, S. Thiruvenkadam, F. Huang, D. Wilson,K. S. Gopinath, R. W. Briggs, and E. A. Geiser. Using Prior Shapesin Geometric Active Contours in a Variational Framework.International Journal of Computer Vision, 50(3):315–328, 2002. !pages 31[25] Y. Chen, F. Huang, H. D. Tagare, and M. Rao. A CoupledMinimization Problem for Medical Image Segmentation with Priors.International Journal of Computer Vision, 71(3):259–272, 2007. !pages 31, 32[26] S. Chow, P. Nair, K. Gin, T. M. Tsang, P. Lee, and J. Jue.Indeterminate Diastolic Function in the Elderly. In Candian Journalof Cardiology, volume 26, pages 94D—-94D, 2010. ! pages 26, 76[27] L. D. Cohen. On active contour models and balloons. CVGIP:Image Understanding, 53(2):211–218, 1991. ! pages 29[28] T. Cootes and C. Taylor. A mixture model for representing shapevariation. Image and Vision Computing, 17(8):567–573, 1999. !pages 32[29] T. Cootes, C. Taylor, D. Cooper, and J. Graham. Active ShapeModels-Their Training and Application. Computer Vision andImage Understanding, 61(1):38–59, 1995. ! pages 4, 32, 3387[30] T. Cootes, G. Edwards, and C. Taylor. Active appearance models.IEEE Transactions on Pattern Analysis and Machine Intelligence,23(6):681–685, 2001. ! pages 33[31] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham. Trainingmodels of shape from sets of examples. In British Machine VisionConference, pages 9–18. Springer, 1992. ! pages 51[32] C. Cortes and V. Vapnik. Support-vector networks. Machinelearning, 20(3):273–297, 1995. ! pages 8[33] A. R. de Alexandria, P. C. Cortez, J. A. Bessa, J. H. da Silva Fe´lix,J. S. a. de Abreu, and V. H. C. de Albuquerque. pSnakes: a newradial active contour model and its application in the segmentationof the left ventricle from echocardiographic images. ComputerMethods and Programs in Biomedicine, 116(3):260–73, 2014. !pages 29[34] L. der Maaten and G. Hinton. Visualizing data using t-SNE.Journal of Machine Learning Research, 9(2579-2605):85, 2008. !pages 9, 68[35] M. R. Di Tullio, R. L. Sacco, R. R. Sciacca, and S. Homma. Leftatrial size and the risk of ischemic stroke in an ethnically mixedpopulation. Stroke, 30(10):2019–24, 1999. ! pages 3, 19[36] L. R. Dice. Measures of the amount of ecologic association betweenspecies. Ecology, 26(3):297–302, 1945. ISSN 0012-9658. ! pages53[37] T. Dietenbeck, M. Alessandrini, D. Barbosa, J. D’Hooge,D. Friboulet, and O. Bernard. Detection of the whole myocardiumin 2D-echocardiography for multiple orientations using ageometrically constrained level-set. Medical Image Analysis, 16(2):386–401, 2012. ! pages 31, 36[38] J. L. Dorosz, D. C. Lezotte, D. A. Weitzenkamp, L. A. Allen, andE. E. Salcedo. Performance of 3-dimensional echocardiography inmeasuring left ventricular volumes and ejection fraction: asystematic review and meta-analysis. Journal of the AmericanCollege of Cardiology, 59(20):1799–808, 2012. ! pages 1388[39] I. Dydenko, F. Jamal, O. Bernard, J. D’hooge, I. E. Magnin, andD. Friboulet. A level set framework with a shape and motion priorfor segmentation and region tracking in echocardiography. MedicalImage Analysis, 10(2):162–77, 2006. ! pages 4, 30[40] B. Georgescu, X. Zhou, D. Comaniciu, and A. Gupta.Database-Guided Segmentation of Anatomical Structures withComplex Appearance. In IEEE Computer Society Conference onComputer Vision and Pattern Recognition, volume 2, pages429–436. IEEE, 2005. ! pages 4, 36[41] L. Grady, S. Datta, O. Kutter, C. Duong, W. Wein, S. Little, S. Igo,S. Liu, and M. Vannan. Regurgitation Quantification Using 3DPISA in Volume Echocardiography. In G. Fichtinger, A. Martel, andT. Peters, editors,Medical Image Computing andComputer-Assisted Intervention MICCAI 2011 SE - 63, volume6893 of Lecture Notes in Computer Science, pages 512–519.Springer Berlin Heidelberg, 2011. ! pages 25[42] G. Hamarneh and T. Gustavsson. Deformable spatio-temporal shapemodels: extending active shape models to 2D+time. Image andVision Computing, 22(6):461–470, 2004. ! pages 33[43] A. K. Hamou and M. R. El-Sakka. Optical Flow Active Contourswith Primitive Shape Priors for Echocardiography. EURASIPJournal on Advances in Signal Processing, 2010:1–11, 2010. !pages 31[44] B. D. Hoit. Assessment of echocardiographic left atrial size: howaccurate do we need to be? Journal of American College ofCardiology Cardiovascular Imaging, 5(8):778–80, 2012. ! pages66[45] R. I. Ionasec, I. Voigt, B. Georgescu, Y. Wang, H. Houle,F. Vega-Higuera, N. Navab, and D. Comaniciu. Patient-specificmodeling and quantification of the aortic and mitral valves from 4-Dcardiac CT and TEE. IEEE Transactions on Medical Imaging, 29(9):1636–51, 2010. ! pages 25[46] G. C. Kane, B. L. Karon, D. W. Mahoney, M. M. Redfield, V. L.Roger, J. C. Burnett, S. J. Jacobsen, and R. J. Rodeheffer.89Progression of left ventricular diastolic dysfunction and risk of heartfailure. Journal of the American Medical Association, 306(8):856–63, 2011. ! pages 26[47] M. Kass, A. Witkin, and D. Terzopoulos. Snakes: Active contourmodels. International Journal of Computer Vision, 1(4):321–331,1988. ! pages 28[48] E. K. Kerut, E. F. McIlwain, and G. D. Plotnick. Handbook ofEcho-Doppler Interpretation. John Wiley & Sons, 2008. ! pages68, 82[49] B. Kircher, J. A. Abbott, S. Pau, R. G. Gould, R. B. Himelman,C. B. Higgins, M. J. Lipton, and N. B. Schiller. Left atrial volumedetermination by biplane two-dimensional echocardiography:Validation by cine computed tomography. American Heart Journal,121(3):864–871, 1991. ! pages 20[50] J. W. Klingler, C. L. Vaughan, T. D. Fraker, and L. T. Andrews.Segmentation of echocardiographic images using mathematicalmorphology. IEEE Transactions on Biomedical Engineering, 35(11):925–34, 1988. ! pages 4[51] S. Kou, L. Caballero, R. Dulgheru, D. Voilliot, C. De Sousa,G. Kacharava, G. D. Athanassopoulos, D. Barone, M. Baroni,N. Cardim, J. J. Gomez De Diego, A. Hagendorff, C. Henri,K. Hristova, T. Lopez, J. Magne, G. De La Morena, B. A. Popescu,M. Penicka, T. Ozyigit, J. D. Rodrigo Carbonero, A. Salustri, N. VanDe Veire, R. S. Von Bardeleben, D. Vinereanu, J.-U. Voigt, J. L.Zamorano, E. Donal, R. M. Lang, L. P. Badano, and P. Lancellotti.Echocardiographic reference ranges for normal cardiac chambersize: results from the NORRE study. European Heart Journal -Cardiovascular Imaging, 15(6):680–690, 2014. ! pages 21[52] R. M. Lang, M. Bierig, R. B. Devereux, F. A. Flachskampf,E. Foster, P. A. Pellikka, M. H. Picard, M. J. Roman, J. Seward, J. S.Shanewise, S. D. Solomon, K. T. Spencer, M. S. J. Sutton, and W. J.Stewart. Recommendations for chamber quantification: a reportfrom the American Society of Echocardiography’s Guidelines andStandards Committee and the Chamber Quantification WritingGroup, developed in conjunction with the European Association of90Echocardiograph. Journal of the American Society ofEchocardiography, 18(12):1440–63, 2005. ! pages 21[53] R. M. Lang, L. P. Badano, V. Mor-Avi, J. Afilalo, A. Armstrong,L. Ernande, F. A. Flachskampf, E. Foster, S. A. Goldstein,T. Kuznetsova, P. Lancellotti, D. Muraru, M. H. Picard, E. R.Rietzschel, L. Rudski, K. T. Spencer, W. Tsang, and J.-U. Voigt.Recommendations for Cardiac Chamber Quantification byEchocardiography in Adults: An Update from the American Societyof Echocardiography and the European Association ofCardiovascular Imaging. Journal of the American Society ofEchocardiography, 28(1):1–39.e14, 2015. ! pages 3, 20, 21, 54[54] S. J. Lester, E. W. Ryan, N. B. Schiller, and E. Foster. Best methodin clinical practice and in research studies to determine left atrialsize. The American Journal of Cardiology, 84(7):829–832, 1999. !pages 20[55] K. Y. E. Leung, M. G. Danilouchkine, M. van Stralen, N. de Jong,A. F. W. van der Steen, and J. G. Bosch. Probabilistic framework fortracking in artifact-prone 3D echocardiograms. Medical ImageAnalysis, 14(6):750–8, 2010. ! pages 36[56] M. Leventon, W. Grimson, and O. Faugeras. Statistical shapeinfluence in geodesic active contours. In Proceedings IEEEConference on Computer Vision and Pattern Recognition. CVPR2000 (Cat. No.PR00662), volume 1, pages 316–323. IEEE Comput.Soc, 2000. ! pages 32[57] Y.-O. Li, T. Adali, and V. D. Calhoun. Estimating the number ofindependent components for functional magnetic resonance imagingdata. Human Brain Mapping, 28(11):1251–66, 2007. ! pages 59[58] N. Lin, W. Yu, and J. S. Duncan. Combinative multi-scale level setframework for echocardiographic image segmentation. MedicalImage Analysis, 7(4):529–537, 2003. ! pages 4, 36[59] J. Lindenfeld, N. M. Albert, J. P. Boehmer, S. P. Collins, J. A.Ezekowitz, M. M. Givertz, S. D. Katz, M. Klapholz, D. K. Moser,J. G. Rogers, R. C. Starling, W. G. Stevenson, W. H. W. Tang, J. R.Teerlink, and M. N. Walsh. HFSA 2010 Comprehensive Heart91Failure Practice Guideline. Journal of Cardiac Failure, 16(6):e1–194, 2010. ! pages 25[60] R. Lozano, M. Naghavi, K. Foreman, S. Lim, K. Shibuya,V. Aboyans, J. Abraham, T. Adair, R. Aggarwal, and S. Y. Ahn.Global and regional mortality from 235 causes of death for 20 agegroups in 1990 and 2010: a systematic analysis for the GlobalBurden of Disease Study 2010. The Lancet, 380(9859):2095–2128,2013. ! pages 3[61] L. Ma, B. Wang, X. Chen, and J. Xiong. Detecting functionalconnectivity in the resting brain: a comparison between ICA andCCA. Magnetic Resonance Imaging Journal, 25(1):47–56, 2007.! pages 59[62] T. Mansi, I. Voigt, E. Mengue, R. Ionasec, B. Georgescu, T. Noack,J. Seeburger, and D. Comaniciu. Towards Patient-SpecificFinite-Element Simulation of MitralClip Procedure. InG. Fichtinger, A. Martel, and T. Peters, editors,Medical ImageComputing and Computer-Assisted Intervention MICCAI 2011 SE -57, volume 6891 of Lecture Notes in Computer Science, pages452–459. Springer Berlin Heidelberg, 2011. ! pages 25[63] J. J. V. McMurray and M. A. Pfeffer. Heart failure. Lancet, 365(9474):1877–89, 2005. ! pages 25[64] M. Mignotte and J. Meunier. A multiscale optimization approachfor the dynamic contour-based boundary detection issue.Computerized Medical Imaging and Graphics, 25(3):265–275,2001. ! pages 4, 29[65] I. Mikic´, S. Krucinski, and J. D. Thomas. Segmentation andtracking in echocardiographic sequences: active contours guided byoptical flow estimates. IEEE Transactions on Medical Imaging, 17(2):274–84, 1998. ! pages 29[66] A. Mishra, P. Dutta, and M. Ghosh. A GA based approach forboundary detection of left ventricle with echocardiographic imagesequences. Image and Vision Computing, 21(11):967–976, 2003. !pages 4, 2992[67] S. C. Mitchell, J. G. Bosch, B. P. F. Lelieveldt, R. J. van der Geest,J. H. C. Reiber, and M. Sonka. 3-D active appearance models:segmentation of cardiac MR and ultrasound images. IEEETransactions on Medical Imaging, 21(9):1167–78, 2002. ! pages33[68] J. E. Moller, G. S. Hillis, J. K. Oh, J. B. Seward, G. S. Reeder, R. S.Wright, S. W. Park, K. R. Bailey, and P. A. Pellikka. Left atrialvolume: a powerful predictor of survival after acute myocardialinfarction. Circulation, 107(17):2207–12, 2003. ! pages 3, 19[69] V. Mor-Avi, C. Yodwut, C. Jenkins, H. Ku¨hl, H.-J. Nesser, T. H.Marwick, A. Franke, L. Weinert, J. Niel,R. Steringer-Mascherbauer, B. H. Freed, L. Sugeng, and R. M.Lang. Real-time 3D echocardiographic quantification of left atrialvolume: multicenter study for validation with CMR. The Journal ofthe American College of Cardiology Cardiovascular imaging, 5(8):769–77, 2012. ! pages 3[70] S. F. Nagueh, C. P. Appleton, T. C. Gillebert, P. N. Marino, J. K. Oh,O. A. Smiseth, A. D. Waggoner, F. A. Flachskampf, P. A. Pellikka,and A. Evangelista. Recommendations for the evaluation of leftventricular diastolic function by echocardiography. Journal of theAmerican Society of Echocardiography, 22(2):107–33, 2009. !pages 26[71] R. A. Nishimura, C. M. Otto, R. O. Bonow, B. A. Carabello, J. P.Erwin, R. A. Guyton, P. T. O’Gara, C. E. Ruiz, N. J. Skubas,P. Sorajja, T. M. Sundt, and J. D. Thomas. 2014 AHA/ACCGuideline for the Management of Patients With Valvular HeartDisease: A Report of the American College ofCardiology/American Heart Association Task Force on PracticeGuidelines. Journal of the American College of Cardiology, 63(22):e57–e185, 2014. ! pages 22, 24[72] J. Noble and D. Boukerroui. Ultrasound image segmentation: asurvey. IEEE Transactions on Medical Imaging, 25(8):987–1010,2006. ! pages 4, 28[73] S. Osher and J. A. Sethian. Fronts propagating withcurvature-dependent speed: Algorithms based on Hamilton-Jacobi93formulations. Journal of Computational Physics, 79(1):12–49,1988. ! pages 29[74] N. Paragios and R. Deriche. Geodesic active regions and level setmethods for motion estimation and tracking. Computer Vision andImage Understanding, 97(3):259–282, 2005. ! pages 33[75] W. J. Paulus, C. Tscho¨pe, J. E. Sanderson, C. Rusconi, F. A.Flachskampf, F. E. Rademakers, P. Marino, O. A. Smiseth, G. DeKeulenaer, A. F. Leite-Moreira, A. Borbe´ly, I. E´des, M. L. Handoko,S. Heymans, N. Pezzali, B. Pieske, K. Dickstein, A. G. Fraser, andD. L. Brutsaert. How to diagnose diastolic heart failure: a consensusstatement on the diagnosis of heart failure with normal leftventricular ejection fraction by the Heart Failure andEchocardiography Associations of the European Society ofCardiology. European Heart Journal, 28(20):2539–2550, 2007. !pages 76[76] H. Peng, F. Long, and C. Ding. Feature selection based on mutualinformation: criteria of max-dependency, max-relevance, andmin-redundancy. IEEE Transactions on Pattern Analysis andMachine Intelligence, 27(8):1226–38, 2005. ! pages 9, 55[77] A. M. Pritchett, S. J. Jacobsen, D. W. Mahoney, R. J. Rodeheffer,K. R. Bailey, and M. M. Redfield. Left atrial volume as an index ofleft atrial size: a population-based study. Journal of the AmericanCollege of Cardiology, 41(6):1036–43, 2003. ! pages 20[78] M. M. Redfield, S. J. Jacobsen, J. C. Burnett, Jr, D. W. Mahoney,K. R. Bailey, and R. J. Rodeheffer. Burden of Systolic and DiastolicVentricular Dysfunction in the Community. Journal of theAmerican Medical Association, 289(2):194, 2003. ! pages 25[79] D. Reed, R. D. Abbott, M. L. Smucker, and S. Kaul. Prediction ofoutcome after mitral valve replacement in patients withsymptomatic chronic mitral regurgitation. The importance of leftatrial size. Circulation, 84(1):23–34, 1991. ! pages 72[80] B. Ristow, S. Ali, M. A. Whooley, and N. B. Schiller. Usefulness ofleft atrial volume index to predict heart failure hospitalization andmortality in ambulatory patients with coronary heart disease and94comparison to left ventricular ejection fraction (from the Heart andSoul Study). The American Journal of Cardiology, 102(1):70–6,2008. ! pages 3, 19[81] M. Rousson and N. Paragios. Shape Priors for Level SetRepresentations. European Conference on Computer Vision, pages78–92, 2002. ! pages 31, 33[82] E. Sakaguchi, A. Yamada, K. Sugimoto, Y. Ito, K. Shiino,K. Takada, M. Iwase, and Y. Ozaki. Prognostic value of left atrialvolume index in patents with first acute myocardial infarction.European Journal of Echocardiography, 12(6):440–4, 2011. !pages 3, 19[83] A. Sarti, C. Corsi, E. Mazzini, and C. Lamberti. Maximumlikelihood segmentation of ultrasound images with Rayleighdistribution. IEEE Transactions on Ultrasonics, Ferroelectrics andFrequency Control, 52(6):947–960, 2005. ! pages 4, 30, 36[84] S. Schabelman, N. B. Schiller, N. H. Silverman, and T. A. Ports.Left atrial volume estimation by two-dimensionalechocardiography. Catheterization and Cardiovascular Diagnosis,7(2):165–178, 1981. ! pages 20[85] R. Schneider, N. Tenenholtz, D. Perrin, G. Marx, P. del Nido, andR. Howe. Patient-Specific Mitral Leaflet Segmentation from 4DUltrasound. In G. Fichtinger, A. Martel, and T. Peters, editors,Medical Image Computing and Computer-Assisted InterventionMICCAI 2011 SE - 64, volume 6893 of Lecture Notes in ComputerScience, pages 520–527. Springer Berlin Heidelberg, 2011. !pages 25[86] C. Sheng, Y. Xin, Y. Liping, and S. Kun. Segmentation inechocardiographic sequences using shape-based snake modelcombined with generalized Hough transformation. TheInternational Journal of Cardiovascular Imaging, 22(1):33–45,2006. ! pages 4[87] Z. Tao, H. D. Tagare, and J. D. Beaty. Evaluation of four probabilitydistribution models for speckle in clinical cardiac ultrasoundimages. IEEE Transactions on Medical Imaging, 25(11):1483–91,2006. ! pages 3495[88] C. Tobon-Gomez, A. Geers, J. Peters, J. Weese, K. Pinto, R. Karim,T. Schaeffter, R. Razavi, and K. Rhode. Benchmark for algorithmssegmenting the left atrium from 3D CT and MRI datasets. IEEETransactions on Medical Imaging, 34(7):1460–1473, 2015. !pages 64[89] B. H. Trichon, G. Felker, L. K. Shaw, C. H. Cabell, and C. M.OConnor. Relation of frequency and severity of mitral regurgitationto survival among patients with left ventricular systolic dysfunctionand heart failure. The American Journal of Cardiology, 91(5):538–543, 2003. ! pages 3[90] A. Tsai, A. Yezzi, W. Wells, C. Tempany, D. Tucker, A. Fan, W. E.Grimson, and A. Willsky. A shape-based approach to thesegmentation of medical imagery using level sets. IEEETransactions on Medical Imaging, 22(2):137–54, 2003. ! pages 32[91] T. S. Tsang, M. E. Barnes, B. J. Gersh, K. R. Bailey, and J. B.Seward. Risks for atrial fibrillation and congestive heart failure inpatients 65 years of age with abnormal left ventricular diastolicrelaxation. The American Journal of Cardiology, 93(1):54–58,2004. ! pages 2[92] T. Utsunomiya, R. Doshi, D. Patel, D. Nguyen, K. Meht, and J. M.Gardin. Regurgitant Volume Estimation in Patients with MitralRegurgitation: Initial Studies Using the Color Doppler ”ProximalIsovelocity Surface Area” Method. Echocardiography, 9(1):63–70,1992. ! pages 23[93] F. Veronesi, C. Corsi, E. Caiani, L. Sugeng, L. Weinert, V. Mor-Avi,R. Lang, and C. Lamberti. Semi-automatic tracking for mitralannulus dynamic analysis using real-time 3D echocardiography.Computers in Cardiology, pages 113–116, 2006. ! pages 25[94] R. Wagner, S. Smith, J. Sandrik, and H. Lopez. Statistics of Specklein Ultrasound B-Scans. IEEE Transactions on Sonics andUltrasonics, 30(3):156–163, 1983. ! pages 29[95] Y. Wang. Atrial volume in a normal adult population bytwo-dimensional echocardiography. CHEST Journal, 86(4):595,1984. ! pages 7396[96] Y. Wang, D. Vitanovski, B. Georgescu, R. Ionasec, I. Voigt,S. Datta, C. Gruner, B. Herzog, P. Biaggi, G. Funka-Lea, andD. Comaniciu. Automatic Detection and Quantification of MitralRegurgitation on TTE with Application to Assist Mitral ClipPlanning and Evaluation. In K. Drechsler, M. Erdt, M. Linguraru,C. Oyarzun Laura, K. Sharma, R. Shekhar, and S. Wesarg, editors,Clinical Image-Based Procedures. From Planning to InterventionSE - 5, volume 7761 of Lecture Notes in Computer Science, pages33–41. Springer Berlin Heidelberg, 2013. ! pages 22, 25, 82[97] WHO. World Health Organization. Global status report onnon-communicable diseases, 2010. ! pages 1[98] I. Wolf, M. Hastenteufel, R. De Simone, M. Vetter, G. Glombitza,S. Mottl-Link, C. F. Vahl, and H.-P. Meinzer. ROPES: asemiautomated segmentation method for accelerated analysis ofthree-dimensional echocardiographic data. IEEE Transactions onMedical Imaging, 21(9):1091–104, 2002. ! pages 36[99] C. Xu and J. L. Prince. Snakes, shapes, and gradient vector flow.IEEE transactions on image processing : a publication of the IEEESignal Processing Society, 7(3):359–69, 1998. ! pages 29[100] J. Yan and T. Zhuang. Applying improved fast marching method toendocardial boundary detection in echocardiographic images.Pattern Recognition Letters, 24(15):2777–2784, 2003. ! pages 29,30[101] Y. Yue and H. D. Tagare. Evaluation of a Level Set SegmentationMethod for Caridac Ultrasound Images. In S. A. McAleavey andJ. D’hooge, editors,Medical Imaging, pages 69200E–69200E–8.International Society for Optics and Photonics, 2008. ! pages 36[102] S. K. Zhou. Shape regression machine and efficient segmentation ofleft ventricle endocardium from 2D B-mode echocardiogram.Medical Image Analysis, 14(4):563–81, 2010. ! pages 4, 33, 36[103] Y. Zhu, X. Papademetris, A. J. Sinusas, and J. S. Duncan. A coupleddeformable model for tracking myocardial borders from real-timeechocardiography using an incompressibility constraint. MedicalImage Analysis, 14(3):429–48, 2010. ! pages 3697[104] M. R. Zile and D. L. Brutsaert. New concepts in diastolicdysfunction and diastolic heart failure: Part I: diagnosis, prognosis,and measurements of diastolic function. Circulation, 105(11):1387–93, 2002. ! pages 25[105] M. R. Zile, C. F. Baicu, and W. H. Gaasch. Diastolic heartfailure–abnormalities in active relaxation and passive stiffness of theleft ventricle. The New England Journal of Medicine, 350(19):1953–9, 2004. ! pages 26[106] M. R. Zile, C. F. Baicu, and D. D. Bonnema. Diastolic HeartFailure: Definitions and Terminology. Progress in CardiovascularDiseases, 47(5):307–313, 2005. ! pages 25[107] M. R. Zile, J. S. Gottdiener, S. J. Hetzel, J. J. McMurray,M. Komajda, R. McKelvie, C. F. Baicu, B. M. Massie, and P. E.Carson. Prevalence and significance of alterations in cardiacstructure and function in patients with heart failure and a preservedejection fraction. Circulation, 124(23):2491–501, 2011. ! pages 3,19[108] W. A. Zoghbi, M. Enriquez-Sarano, E. Foster, P. A. Grayburn, C. D.Kraft, R. A. Levine, P. Nihoyannopoulos, C. M. Otto, M. A.Quinones, H. Rakowski, W. J. Stewart, A. Waggoner, and N. J.Weissman. Recommendations for evaluation of the severity ofnative valvular regurgitation with two-dimensional and Dopplerechocardiography. Journal of the American Society ofEchocardiography, 16(7):777–802, 2003. ! pages 24, 7198


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items