Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Using an ultrasound-derived model to assist in dosimetry for prostate cancer treatment through brachytherapy Tam, Cindy R. 2013

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

Item Metadata


24-ubc_2013_fall_tam_cindy.pdf [ 8.62MB ]
JSON: 24-1.0074132.json
JSON-LD: 24-1.0074132-ld.json
RDF/XML (Pretty): 24-1.0074132-rdf.xml
RDF/JSON: 24-1.0074132-rdf.json
Turtle: 24-1.0074132-turtle.txt
N-Triples: 24-1.0074132-rdf-ntriples.txt
Original Record: 24-1.0074132-source.json
Full Text

Full Text

Using an Ultrasound-Derived Modelto Assist in Dosimetry for ProstateCancer Treatment ThroughBrachytherapybyCindy R. TamB.Sc., University of Victoria, 2001M.Sc., McGill University, 2005A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFMASTER OF SCIENCEinThe Faculty of Graduate Studies(Physics)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)August 2013c? Cindy R. Tam 2013AbstractProstate cancer is the most common form of cancer affecting men in Canada.Patients with localized, early stage disease are often treated using prostatebrachytherapy, a technique that involves surgically implanting small radioac-tive capsules or ?seeds? in the prostate. Implantation is performed underthe guidance of transrectal ultrasound (TRUS) imaging, and treatment isassessed postoperatively for quality assurance purposes. Pelvic CT imagingis used to evaluate the dose delivered to the target; however, it is a chal-lenge to consistently and confidently identify prostate boundaries due to thepoor soft tissue contrast on CT. This leads to large variability in CT-definedanatomical contours and calculated dosimetric quality assurance parameters,and has led to increased reliance on other imaging technologies such as MR.Meanwhile, TRUS typically provides high-quality anatomical visualization,but provides insufficient information for dose calculation purposes.We have developed a new method to transfer ultrasound-based contoursto CT images using mathematical modeling and a novel registration tech-nique. The prostate model, derived from TRUS contours, is generated viatwo streams: one assumes a modified ellipsoid shape (model X), and theother performs a straightforward linear interpolation (model Y). Both aremanipulated to account for expected deformations such as TRUS-probe com-pression and edema. Registration from TRUS to CT spatial coordinatesis based on matched seed locations. We evaluate the quality of model-generated contours primarily by comparing the measured volume and dosi-metric parameters to the observed variability range determined from manualCT contours. In 19 of model X, and 18 of model Y, cases out of 20, volumesproduced were within the variability observed from 5 experienced physicians.However, dose parameters agreed in only a moderate number of cases (9?13),partly motivating a region-specific analysis. We found the least agreementin the posterior apex, with model contours tending to be larger. We discussthe possible reasons for this, as well as implications on the role of modelingin an applied clinical setting. Ultimately, the ultrasound-informed modelshows promise, and has many benefits relative to other methods, such thosebased on CT or MR.iiPrefaceAside from the material described below, the work presented in this thesisis original and unpublished.A portion of the modeling algorithm described in Chapter 2 was previ-ously published in the MSc thesis ?Prostate segmentation in ultrasound im-ages using image warping and ellipsoid fitting?, UBC, 2007 by S. Badiei, andthe article ?Semi-automatic segmentation for prostate interventions?, MedImage Anal, 15(2):226?237, 2011, principal investigator Dr. S. Mahdavi.The algorithm behind registration of matched seed locations was adaptedfrom material presented in the article ?Prostate brachytherapy postimplantdosimetry: automatic plan reconstruction of stranded implants?, Med Phys,38(1):327?342, 2011, principal investigator Dr. N. Chng. Other parts of thecomputer code used in the analysis were adapted from existing but unpub-lished software written by N. Chng, R. Kosztyla, and J. Lobo.This research was approved by the University of British Columbia?British Columbia Cancer Agency Research Ethics Board certificate numberH10-03169, titled ?Using ultrasound-based prostate modeling towards a newmethod of prostate brachytherapy quality assessment?.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Abbreviations and Terminology . . . . . . . . . . . . . . xiiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Prostate Cancer . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Management of Prostate Cancer . . . . . . . . . . . . . . . . 31.2.1 Prostatectomy . . . . . . . . . . . . . . . . . . . . . . 31.2.2 Radiation therapy . . . . . . . . . . . . . . . . . . . . 31.2.3 Hormone therapy . . . . . . . . . . . . . . . . . . . . 51.3 Prostate Brachytherapy at BCCA . . . . . . . . . . . . . . . 51.3.1 Treatment planning . . . . . . . . . . . . . . . . . . . 61.3.2 Implantation . . . . . . . . . . . . . . . . . . . . . . . 81.3.3 Postimplant dosimetry . . . . . . . . . . . . . . . . . 101.4 Thesis Motivation . . . . . . . . . . . . . . . . . . . . . . . . 122 Methods and Materials . . . . . . . . . . . . . . . . . . . . . . 162.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.1.1 Collection . . . . . . . . . . . . . . . . . . . . . . . . 162.1.2 Procedure . . . . . . . . . . . . . . . . . . . . . . . . 172.1.3 Software . . . . . . . . . . . . . . . . . . . . . . . . . 172.2 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.2.1 Prostate segmentation in ultrasound . . . . . . . . . . 18ivTable of Contents2.2.2 Seed segmentation in fluoroscopy . . . . . . . . . . . 192.2.3 Seed segmentation in ultrasound . . . . . . . . . . . . 192.2.4 Seed segmentation in CT . . . . . . . . . . . . . . . . 212.3 Model Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 212.3.1 Unwarping . . . . . . . . . . . . . . . . . . . . . . . . 232.3.2 Model X: Untapering and ellipse fitting . . . . . . . . 232.3.3 Model X: 3-D tapered ellipsoid fitting . . . . . . . . . 242.3.4 Model X: Slicing and tapering . . . . . . . . . . . . . 242.3.5 Model Y: Linear interpolation . . . . . . . . . . . . . 252.3.6 Registration and edema . . . . . . . . . . . . . . . . . 252.3.7 CT contour slicing . . . . . . . . . . . . . . . . . . . . 252.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.4.1 STAPLE . . . . . . . . . . . . . . . . . . . . . . . . . 262.4.2 Volumetric and dosimetric analysis . . . . . . . . . . 282.4.3 Statistical testing . . . . . . . . . . . . . . . . . . . . 303 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.1 STAPLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343.2 Volumetric and Dosimetric Analysis . . . . . . . . . . . . . . 383.2.1 Total prostate dosimetry . . . . . . . . . . . . . . . . 383.2.2 Sector dosimetry . . . . . . . . . . . . . . . . . . . . . 383.3 Statistical Testing . . . . . . . . . . . . . . . . . . . . . . . . 404 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.1 Modeling as an Alternative to MR-CT Fusion . . . . . . . . 494.2 Seed Localization in TRUS . . . . . . . . . . . . . . . . . . . 505 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56AppendixA Data Tables and Figures . . . . . . . . . . . . . . . . . . . . . 67vList of Tables2.1 Summary of statistical terminology and the general relation-ship between parameters. . . . . . . . . . . . . . . . . . . . . 282.2 Arrangement of raw data for the Friedman Test. . . . . . . . 313.1 Results of TRUS and fluoroscopy seed cloud reconstructionand comparison. Also indicated under column marked ?Thin?are ThinStrand implant cases. Ntotal is the total number ofseeds known to be present during intraoperative data col-lection, and NTRUS and Nfluoro are the number of TRUSand fluoroscopy seeds successfully reconstructed, respectively.The mean offset between registered, matched seeds is calcu-lated for Ncorr number of seeds for which a correspondencewas found. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.2 Number of cases for which the model STAPLE parameter waswithin the 95% CI variability range of manual values, out ofa maximum 20. . . . . . . . . . . . . . . . . . . . . . . . . . . 383.3 Number of cases for which the model total volume and doseparameter was within the 95% CI variability of manual values,out of a maximum 20. . . . . . . . . . . . . . . . . . . . . . . 383.4 Numbers of cases for which the model sector volume and doseparameter was within the 95% CI variability of manual values,out of a maximum 20. . . . . . . . . . . . . . . . . . . . . . . 403.5 Results of statistical tests. . . . . . . . . . . . . . . . . . . . . 413.6 Absolute differences in the rank sums for the whole prostatevolume between all pairs of observers. Note that values aresymmetric across the main diagonal. Values in bold are greaterthan Fisher?s LSD critical value 19.3244 for volume; valuesmarked with an asterisk (*) are greater than Nemenyi?s crit-ical value 40.2860. . . . . . . . . . . . . . . . . . . . . . . . . 42viList of Tables3.7 Absolute differences in the rank sums for total prostate V100parameter between all pairs of observers. Refer to Table 3.6caption for a description; here, Fischer?s LSD critical value is25.5664. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433.8 Absolute differences in the rank sums for total prostate D90parameter between all pairs of observers. Refer to Table 3.6caption for a description; here, Fischer?s LSD critical value is24.4543. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43A.1 Prostate volume (cc) measured for all 20 patients, based on5 manual (A?E) and 2 model (X and Y) sets of contours. . . 67A.2 Prostate V100 (%) calculated for all 20 patients, based on 5manual (A?E) and 2 model (X and Y) sets of contours. . . . 68A.3 Prostate D90 (Gy) calculated for all 20 patients, based on 5manual (A?E) and 2 model (X and Y) sets of contours. . . . 69A.4 Ranks R(Xij) and rank sums Rj of prostate volume given inTable A.1, for use in the Friedman statistical test and post-hoc analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . 70A.5 Ranks R(Xij) and rank sums Rj of prostate V100 given inTable A.2, for use in the Friedman statistical test and post-hoc analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . 71A.6 Ranks R(Xij) and rank sums Rj of prostate D90 given inTable A.3, for use in the Friedman statistical test and post-hoc analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . 72viiList of Figures1.1 The prostate and nearby organs, as viewed on a sagittal crosssection. Source: National Cancer Institute [34]. Creator:Alan Hoofring. . . . . . . . . . . . . . . . . . . . . . . . . . . 21.2 Example set of TRUS images from a preimplant volume study.Nine axial slices separated by 0.5 cm are shown, beginningabove the prostate base (upper left), and moving inferiorly(left to right, then top to bottom) to below the apex (lowerright). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3 The same example set of TRUS images as in Figure 1.2, withcontours overlaid indicating the CTV (red) and PTV (cyan). 81.4 Example treatment plan, overlaid on the same set of TRUSimages as in Figure 1.3. Also indicated are the implant tem-plate grid (yellow crosses), needle locations (open yellow cir-cles), seed locations (filled cyan circles), and the 100% (green),150% (orange), and 200% (magenta) isodose contours. . . . . 91.5 Prostate brachytherapy being performed under ultrasoundguidance. Source: Mayo Foundation for Medical Educationand Research [25]. Used with permission from rights reserved. . . . . . . . . . . . . . . . . . . . . . . . . 101.6 Example set of RO-defined contours overlaid on axial CT im-ages (left) and rendered as 3-D surfaces (right). Anatomicalstructures included are the prostate (red), seminal vesicles(orange), bladder (yellow), urethra (yellow-green), and rec-tum (blue). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111.7 An example demonstrating the large interobserver variabilitythat can exist in CT contouring. Shown here are an axialCT scan slice located towards the prostate base without (topleft) and with (top right) contours overlaid. Contours weremanually defined by 5 experienced physicians. We also showthe same contours as seen on a sagittal (bottom left) andcoronal (bottom right) representation. . . . . . . . . . . . . . 13viiiList of Figures1.8 Midgland prostate slices as seen on intraoperative TRUS (left)and postoperative CT (right). Images are approximately scaledto the same dimensions. The intraoperative TRUS image wasobtained after 3 strands had been implanted, while the post-operative CT was obtained after 22. Note how soft tissueboundaries are distinct on TRUS, but seeds locations are ob-scured, while the opposite is true on CT. . . . . . . . . . . . . 152.1 Sagittal ultrasound images, without (left) and with (right) anRO-defined contour overlaid. . . . . . . . . . . . . . . . . . . 192.2 A fluoroscopy coronal projection, containing 22 seeds. TheTRUS probe and Foley catheter are also visible. . . . . . . . . 202.3 Flowchart illustrating the two modeling streams: the taperedellipsoid model X, and linear interpolation model Y. We haveemphasized in bold the modeling steps that are unique to thisproject. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.4 Schematic diagram illustrating the 9 sector divisions. . . . . . 292.5 Schematic showing the 9 sectors labeled as visualized on asagittal view of the prostate. . . . . . . . . . . . . . . . . . . 303.1 Model X (left) and Y (right) final products are shown ascolour surfaces, with the CT contours overlaid in colouredlines. Matched CT (solid squares) and ultrasound seeds (opensquares) are indicated by a solid line connecting them. Notethat extraprostatic seeds are intentionally placed in the pos-terior to achieve the treatment planning goals of delivering ahigh dose in the posterior peripheral zone. . . . . . . . . . . . 333.2 An example set of evaluated STAPLE weight parameter Wimaps. Note that voxel values tend to be nearly 0 or 1, so theconsensus volume is largely independent of our choice of Withreshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363.3 Results of STAPLE analysis. Mean sensitivity p and positivepredictive values PV(1) and 95% CIs derived from manualcontours are shown (blue crosses and error bars), along withmodel X (red circles) and Y (green squares) values. . . . . . . 37ixList of Figures3.4 Example illustrating high sensitivity with low positive pre-dictive value. Shown are contours generated by three manual?decisions? DA, DB, and DC , the consensus region represent-ing the estimated true segmentation T , and one model de-cision DX . In this scenario, all decision contours result inroughly the same sensitivity, but by selecting a proportion-ally larger area outside of T , DX will have a relatively smallerpositive predictive value. . . . . . . . . . . . . . . . . . . . . . 393.5 Volume and dosimetric parameter results for total prostate.Mean manual contouring values and 95% CIs (blue crossesand error bars) are shown along with values produced bymodel X (red circles) and model Y (green squares). Overplot-ted are the dosimetric ?sub-optimal? QA thresholds (dottedlines): V100 = 85% and D90 = 130 Gy and 180 Gy. . . . . . . 443.6 Colour sector diagrams representing the level of agreement involume, V100 and D90 for models X and Y. The colour scaleon the right indicates the number of cases, out of a maximum20, for which the model parameter was within the 95% CIvariability of manual values, in each sector as seen from thesagittal view (refer to Figure 2.5). . . . . . . . . . . . . . . . 453.7 Sagittal, axial and coronal view of a patient?s CT scan, withSTAPLE (cyan) and model X (magenta) contours overlaid.Seed locations are also marked by cyan bars. The apical pos-terior sector (APS) tends to be larger on model-generatedcontours than it is typically defined manually. . . . . . . . . . 46A.1 Volumetric and dosimetric results for the BAS. Mean man-ual contouring values and 95% CIs (blue crosses and errorbars) are shown along with values derived from model X(red circles) and Y (green squares) contours. Overplotted arethe dosimetric ?sub-optimal? QA thresholds (dotted lines):V100= 85% and D90= 130 Gy and 180 Gy. . . . . . . . . . . . 73A.2 Volumetric and dosimetric results for the MAS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 74A.3 Volumetric and dosimetric results for the AAS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 75A.4 Volumetric and dosimetric results for the BLS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 76A.5 Volumetric and dosimetric results for the MLS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 77xList of FiguresA.6 Volumetric and dosimetric results for the ALS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 78A.7 Volumetric and dosimetric results for the BPS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 79A.8 Volumetric and dosimetric results for the MPS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 80A.9 Volumetric and dosimetric results for the APS. Refer to Fig-ure A.1 caption for a description. . . . . . . . . . . . . . . . . 81xiList of Abbreviations andTerminology2-D/3-D 2-/3-DimensionalAAPM American Association of Physicists in MedicineAAS Apex Anterior SectorADT Androgen Deprivation TherapyALS Apex Lateral SectorAPS Apex Posterior SectorASQ Anterior Superior QuadrantBAS Base Anterior SectorBCCA British Columbia Cancer AgencyBLS Base Lateral SectorBPS Base Posterior SectorCT Computed TomographyCTV Clinical Target VolumeD90 The minimum dose delivered to the hottest 90% of the volumeDRE Digital Rectal ExamDVH Dose Volume HistogramGy GrayHDR High Dose RatexiiList of Abbreviations and TerminologyLDR Low Dose RateMAS Midgland Anterior SectorMLS Midgland Lateral SectorMPS Midgland Posterior SectorMR Magnetic ResonancePSA Prostate Specific AntigenPTV Planning Target VolumeQA Quality AssuranceRO Radiation OncologistSTAPLE Simultaneous Truth and Performance Level EstimationTG Task GroupTRUS Transrectal UltrasoundV 100 The volume receiving at least 100% of the prescription dosexiiiAcknowledgmentsFirst and foremost, I would like to thank my supervisor Ingrid Spadinger,whose wealth of knowledge, guidance, and support were invaluable to me andthe success of my research. Many thanks are also due to my co-supervisorSteven Thomas and departmental supervisor Alex MacKay, for their unceas-ing encouragement. I owe special thanks to Nick Chng, whose expertise inprogramming was indispensable to my project, and who seemed to know theanswer to just about everything I had to ask about anything. Special thanksalso to Jim Morris, who donated much of his valuable time and expertise tothis work. I am extremely grateful to the following physicians at the BCCAwho volunteered to participate in my study and kindly donated their skillsand time: Muneeza Akbar, Arthur Cheung, Juanita Crook, Mira Keyes,Mitchell Liu, Michael McKenzie, Jim Morris, and Tom Pickles. Many othershave generously contributed valuable advice and assistance, including SaraMahdavi, Rob Kosztyla, Julio Lobo, Mehdi Moradi and Tim Salcudean. Iam indebted to BCCA staff who were instrumental during the collection ofmy data, especially radiation therapists, who took on responsibilities I neverasked of them, and made my life easier in the process.Finally, I thank my family for their constant support, especially myhusband Aaron, and our son Elijah, whose immeasurable contribution wassimply his arrival in the middle of all this.xivChapter 1Introduction1.1 Prostate CancerProstate cancer is the most commonly diagnosed form of cancer affectingmen in Canada, accounting for approximately 25% of new cancer diagnosesso far in 2013 [12]. The prostate, a male reproductive gland, secretes partof the fluid in semen. It is located inferior to the bladder and anterior tothe rectum, with the urethra passing through it, as seen in Figure 1.1. Alsosuperior and posterior to the gland are the seminal vesicles. The androgentestosterone stimulates the reproduction and growth of prostate tissue cells,including both normal and cancerous cells. Symptoms of prostate cancermay include urinary or erectile dysfunction; however, since it is common forthe prostate to naturally become enlarged as men age, these symptoms donot necessarily indicate the presence of cancer. Conversely, many men withprostate cancer are asymptomatic. The typical prostate gland in healthyyoung adult men is roughly the size and shape of a walnut: approximately3 by 4 by 2 cm in dimension (?20?30 cc), and weighing ?20 g [45]. Thesuperior-most aspect of the gland is commonly referred to as the base, andthe inferior-most as the apex; this terminology will be used throughout thiswork. The prostate can generally be described by 4 main regions: the centralzone, peripheral zone, transitional zone, and anterior fibromuscular stroma[47]. Prostate cancers are usually adenocarcinomas, occurring in epithelialcells of glandular tissue, and most commonly, although not exclusively, in-volve the peripheral zone. The primary risk factor for prostate cancer isage, increasing for men over 50; other risk factors include family history andrace, and possibly diet and lifestyle.Testing for prostate cancer typically involves a physical exam by a physi-cian (a digital rectal exam, DRE) and a blood test for elevated levels ofprostate specific antigen (PSA). Normal PSA levels are considered less than?4.0 ng/mL, although this value is highly dependent on age [76], and a highPSA level alone is not necessarily indicative of disease. Whether regularPSA screening for men ought to be recommended is a subject of continuingcontroversy [28]. Diagnosis is usually made through pathological testing of11.1. Prostate CancerFigure 1.1: The prostate and nearby organs, as viewed on a sagittal crosssection. Source: National Cancer Institute [34]. Creator: Alan Hoofring.biopsied material, obtained from ?6?12 core samples using needle extractionunder the guidance of transrectal ultrasound (TRUS) imaging. If present,a clinical stage is determined according to the TNM classification system,which describes the site and size of the primary tumour (T1?T4), nodal in-volvement (N0 or N1) and metastatic spread (M0 or M1). Also, a Gleasonscore of 2?10 is calculated from the grade assigned to the biopsy material,which describes the level of cell differentiation and indicates how aggressivea tumour may be. Patients are stratified into low, intermediate or high riskgroups based on tumour stage, PSA level, and Gleason score [2, 20].21.2. Management of Prostate Cancer1.2 Management of Prostate CancerOnce prostate cancer has been diagnosed, the most common forms of man-agement are active surveillance, or radical treatment with the intent to curethrough prostatectomy and/or radiation therapy. Prostate cancer is a rel-atively slowly progressing disease with high overall survival rates. For thisreason, active surveillance is often recommended for low risk patients be-cause it avoids, or at least delays, the side effects of intervention and pre-vents unnecessary over-treatment of patients with favourable risk disease[36]. Under active surveillance, patients are routinely monitored for diseaseprogression through recurrent PSA tests, DREs, and biopsy, and regularlyassessed for continuation of this form of management.1.2.1 ProstatectomyRadical prostatectomy is the surgical removal of the prostate gland, seminalvesicles, and other surrounding tissue. Low to intermediate risk patientswith localized disease (ie. tumour is confined to the prostate) who have?10 years life expectancy are generally eligible for this treatment method,provided they are deemed fit for surgery under anaesthetic. High risk pa-tients may also be eligible for surgery in combination with other forms oftreatment. Open surgery, with the incision in the lower abdomen, is themost common surgical procedure, although surgery may also be performedlaparoscopically. At the beginning of the operation, pelvic lymph nodesare dissected and assessed for microscopic invasion; the outcome of pathol-ogy testing determines whether the procedure should be carried out, andif so, how much tissue should be removed. Potential side effects of radi-cal prostatectomy, beyond those related to the trauma of surgery such aspain or bleeding, include temporary urinary incontinence typically lasting6?12 months (patients are typically catheterized for ?1?2 weeks followingsurgery), erectile dysfunction, although in some cases nerve sparing surgeryto retain sexual function may be an option, and permanent infertility.1.2.2 Radiation therapyRadiation therapy, or radiotherapy, uses high energy ionizing radiation (X-rays or ?-rays) to kill cancer cells by causing lethal DNA strand breaks.However, radiation can damage normal cells as well: thus, the primary goalof radiotherapy is to deliver a sufficiently high dose to the tumour whileinflicting the least amount of damage possible to healthy tissue. The main31.2. Management of Prostate Cancerapproaches to achieving this balance involve limiting the amount of normaltissue exposed in the targeted volume, and utilizing the differing radiobio-logical characteristics of cancerous and normal cells [75]. Fractionation, thetechnique of delivering small doses over a protracted period of time, ampli-fies both the survival abilities of normal tissue and the radiosensitivity oftumour tissue [17, 33]. For prostate cancer, radiation therapy is deliveredthrough one of two ways: external beam radiation, or brachytherapy.External beam radiation is the most common form of radiotherapy forcancer treatment in general, utilizing radiation sourced from outside thebody, typically generated by a linear particle accelerator known as a linac,that is directed to a localized area inside the patient. The complexity of atreatment delivered through external beam can range from 2 or 4 rectangular?beams? directed to the general area of the patient?s tumour, to treatmentplans that involve moving the linac unit and beam-shaping collimators whilethe beam is on, with the goal of irradiating a precise, complex volume thattightly conforms to the shape of the target. Typically, patients receive treat-ment in 20 to 40 fractions over the course of 4 to 8 weeks. Prostate cancerpatients with stage T1?T3 disease are generally eligible for external beamtherapy, although usually only in combination with other forms of treatmentfor higher stage patients. For advanced stage patients (T4), radiation maybe used to relieve symptoms of disease (ie. palliative, rather than radical,therapy). External beam radiation is also used in combination with (adju-vant) or after (salvage therapy) prostatectomy or brachytherapy. Generalside effects of radiation therapy may include fatigue, weight loss, or loss ofappetite, and more specific to the region are short-term urinary and bowelirritation or dysfunction, normally lasting weeks to months, and erectiledysfunction that may be temporary or permanent.Brachytherapy, from the Greek prefix meaning ?short distance?, deliverstreatment through radioactive sources that are implanted directly into thepatient at or near the tumour site. Historically, the first uses of radiationtherapy were brachytherapy, following the discovery of radium in the late19th century [17, 74]. The primary advantage of this technique is that ahigh dose can be deposited close to the target location, with the exposurerapidly falling off with increasing distance from the source, so that nearbystructures may be spared. Because of its highly localized nature, patientsmust have disease that is confined to the prostate (ie. low or intermediaterisk, stages T1?T2c), to be eligible for brachytherapy. In addition, prostateanatomy must be within set size constraints so as to avoid pubic arch in-terference during source implantation, and patients must be suitable candi-dates for surgery under anesthesia. Brachytherapy treatments are typically41.3. Prostate Brachytherapy at BCCAcategorized as being either low dose rate (LDR) or high dose rate (HDR),which describes the radioactive source being used. For HDR treatments,the method of delivery is temporary, in which catheters are used to placethe radioactive source at specific locations for limited amount of time. LDRis most frequently delivered through permanent implantation of radioactivesources. For prostate brachytherapy, permanent implant LDR treatment ismore common, although HDR is becoming more widely used. Permanentimplant brachytherapy is an out-patient procedure requiring an operationunder anesthesia to implant the sources. Side effects of brachytherapy areshort term discomfort during the recovery from surgery, and radiation tox-icity symptoms similar to external beam but reportedly less severe, exceptfor urinary symptoms which tend to be more severe and longer lasting (3?12months or more) [1, 61].1.2.3 Hormone therapyIn some cases, hormone therapy known as androgen deprivation therapy(ADT), which blocks either the production or effect of androgens on prostatecells, may be considered. It is used in conjunction with other treatmenttypes, or alone as a management option for advanced or recurrent cancer.The goal is to relieve symptoms such as prostate swelling, and to slow orreverse tumour growth. For some patients, ADT may also improve theeffectiveness of radiation therapy, and in these cases ADT is administeredfor 3?6 months before, during, and after treatment for up to 3 years (ormore) [61].1.3 Prostate Brachytherapy at BCCAThe Provincial Prostate Brachytherapy Program at the British ColumbiaCancer Agency (BCCA) was founded in 1997, and since then, over 4000LDR implants have been performed [51]. Initially, only low-risk or ?low-tier?intermediate risk patients were treated; since 2009, prostate brachytherapyhas become standard treatment for all low and immediate risk patients, withsome receiving hormone therapy before and after their implant [35]. Currenteligibility criteria is clinical stage ? T2c, PSA ? 20 ng/mL, and Gleasonscore ? 7. Recently, the disease-free survival rates from this program basedon the first 1006 consecutive patients enrolled were reported to be 96.7% for5-years and 94.1% 10-years [51]. Certain high risk patients are eligible toreceive LDR brachytherapy in combination with external beam and hormone51.3. Prostate Brachytherapy at BCCAtherapy. Furthermore, at the BCCA Centre for the Southern Interior, HDRprostate brachytherapy is offered as part of a clinical trial.General recommendations on treatment and quality assurance proce-dures for permanent implant prostate brachytherapy are described in the re-ports of The American Association of Physicists in Medicine (AAPM) TaskGroups #64 [94] and #137 [58]. Various other professional organizationshave published guidelines on performing and analysing permanent prostatebrachytherapy: The American Brachytherapy Society (ABS) [21, 52, 53],The American Society of Radiation Oncology (ASTRO) and American Col-lege of Radiology (ACR) [69]. Here, we outline the standard clinical proce-dures followed in the BCCA program.1.3.1 Treatment planningProstate brachytherapy at the BCCA is carried out following the ?Seattlemethod? [7], based on a technique first reported by Holm et al. in 1993 [32],for which treatment is a 2-step procedure. First is the treatment planningstep, wherein patients undergo a transrectal ultrasound (TRUS) imagingvolume study: a series of parallel axial image ?slices? captured at 0.5 cmintervals that encompass the entire prostate anatomy from above the base tobelow the apex, see Figure 1.2. On this volume study, a preliminary set of or-gan delineations or ?contours? is defined. Contours are generated at BCCAusing in-house semi-automatic contouring software, which creates contoursfor the clinical target volume (CTV). The CTV includes the prostate and asmall portion of the seminal vesicles. These are reviewed by the radiationoncologist (RO), and modified if necessary. Once approved, the planningtarget volume (PTV) is generated, which includes the CTV plus margins,that are added mainly to account for targeting uncertainties; see Figure 1.3.The medical physicist manually generates a preoperative implant plan(or ?preplan?), in which the placement of brachytherapy sources to ob-tain the optimum treatment is determined. For all patients, 144 Gy as aminimum peripheral dose is prescribed to cover at least 98% of the PTV[50], where 1 Gray (Gy) is the SI unit of absorbed dose (equivalent to 1J/kg). The same radioactive isotope?Iodine-125 (125I), which has a half-life of 59.4 days and generates photons of energy ?30 keV [59]?is used forall prostate brachytherapy treatments. The treatment planning algorithmaims to achieve an inverted horseshoe-shaped continuous region of high dose(150%, or 216 Gy) by increasing the source density in the posterior periph-eral zone, while maintaining a lower maximum dose around the urethra;see Figure 1.4. To make the planning and implantation procedures sim-61.3. Prostate Brachytherapy at BCCAFigure 1.2: Example set of TRUS images from a preimplant volume study.Nine axial slices separated by 0.5 cm are shown, beginning above the prostatebase (upper left), and moving inferiorly (left to right, then top to bottom)to below the apex (lower right).pler, faster, and less prone to errors, source positions are planned accordingto a regular template grid pattern, and are symmetric with respect to themidsagittal plane. Other techniques to improve efficiency and reduce errorsinvolve minimizing the number of needle insertions required by maximiz-ing the number of sources per needle, and using uniformly spaced, strandedsources as much as possible (needles and strands are described in ?1.3.2).Dose calculation is performed following the protocols outlined in theAAPM Task Group #43 (TG-43) reports [56, 65]. These documents pro-vide recommendations on the calculation of the 3-dimensional (3-D) dose-rate distribution to water as a function of distance and orientation fromlow energy photon-emitting brachytherapy sources, taking into account thesource strength, geometry, and the effects of attenuation and scatter due tothe source material itself. Once a treatment plan has been approved by themedical physicist and RO, the sources are ordered, and arrive pre-loaded inneedles for the individualized treatment.71.3. Prostate Brachytherapy at BCCAFigure 1.3: The same example set of TRUS images as in Figure 1.2, withcontours overlaid indicating the CTV (red) and PTV (cyan).1.3.2 ImplantationThe second step in the treatment procedure is the brachytherapy sourceimplant itself. Typically, this takes place 4?6 weeks following the date ofthe TRUS volume study. Implantation is performed by the RO, and as-sisted by radiation therapists and nurses, in a surgical procedure with thepatient under anesthesia and placed in lithotomy position. Patients are usu-ally catheterized with a Foley catheter and contrast medium placed in theurethra, for imaging purposes.Individual brachytherapy sources are in the form of small ?seeds? (termswill be used interchangeably throughout this work), approximately the sizeof a rice grain, spaced together at fixed, predetermined center-to-center dis-tances (usually 1 cm) using tissue absorbable spacers. Seeds and spacersare encased in a braided, tissue absorbable sleeve (the ?strand?) that helpsmaintain the correct seed spacing after implantation. Strands are sterilizedprior to the procedure, and implanted through the patient?s perineum usinghollow needles that are inserted using a template grid for guidance, following81.3. Prostate Brachytherapy at BCCAFigure 1.4: Example treatment plan, overlaid on the same set of TRUSimages as in Figure 1.3. Also indicated are the implant template grid (yellowcrosses), needle locations (open yellow circles), seed locations (filled cyancircles), and the 100% (green), 150% (orange), and 200% (magenta) isodosecontours.the same pattern as used in planning. Strand configurations (ie. numberof seeds and seed spacings) vary from needle to needle and are based onindividualized treatment plans. Typically, 80?150 seeds on 18?30 strandsare implanted per treatment.Real-time TRUS imaging is employed during patient set up to ensurethat anatomical information matches between planning and implant data,and also to provide for guidance during needle insertions (see Figure 1.5).X-ray fluoroscopic imaging, from which coronal projections are obtained, isalso utilized for guidance [94]. The two imaging modalities are complimen-tary: sources can be easily localized on X-ray fluoroscopy, while soft tissueanatomy is better visualized on TRUS. The procedure typically requires ?1hour depending on the number of needles required, and patients are releasedfrom the clinic on the same day.91.3. Prostate Brachytherapy at BCCAFigure 1.5: Prostate brachytherapy being performed under ultrasound guid-ance. Source: Mayo Foundation for Medical Education and Research [25].Used with permission from All rights reserved.1.3.3 Postimplant dosimetryEach treatment undergoes quality assurance (QA) as part of routine proce-dures in the prostate brachytherapy program. QA is performed not only todetermine the quality of individual treatments, but to provide feedback forimprovement of future treatments and evaluate the quality and outcomesof the program as a whole [53]. Postimplant dosimetry, which is the mea-surement of dose received by regions of interest (again following the AAPMTG-43 protocol), is one component of QA, and is based on CT (computedtomography) imaging. Patients receive a pelvic CT scan in supine posi-tion following their implantation. At BCCA, this is currently done on thesame day as the implant, and is referred to as Day-0 CT. As a side note,general recommendations state that dosimetric evaluation should ideally beperformed ?4 weeks postimplant, when the effects of prostatic edema arereduced [94]; however, the choice to obtain implant-day CT is based on101.3. Prostate Brachytherapy at BCCAFigure 1.6: Example set of RO-defined contours overlaid on axial CT images(left) and rendered as 3-D surfaces (right). Anatomical structures includedare the prostate (red), seminal vesicles (orange), bladder (yellow), urethra(yellow-green), and rectum (blue).practical considerations for patient accessibility, as well as the benefits ofreceiving immediate feedback on treatment quality [57]. Manual contouringof patient anatomy, including the prostate as well as nearby structures suchas the bladder, urethra, seminal vesicles, and rectum, is performed on Day-0 CT by the RO. Accurate 3-D source positions are localized on CT usingcommercial software; more recently, in-house software that also performs a?plan reconstruction?, wherein individual sources are identified and assigneda strand and seed number according to the treatment preplan [15], has beenadopted into routine procedures.Treatment quality is primarily assessed through dosimetric parameters,which are determined from the quantitative dose distribution graph knownas a dose volume histogram (DVH). On a DVH, the differential or cumulativedose to a target or organ at risk can be plotted for any volumetric structure.As related to the target volume (ie. the prostate), the most importantmetrics utilized in postimplant dosimetry that can be retrieved from theDVH are? V100: the volume receiving at least 100% of the prescription dose,reported in cubic centimeters (cc) or as a percentage of total volume,and? D90: the minimum dose delivered to the hottest 90% of the volume,reported in Gy or as a percentage of prescription dose.111.4. Thesis MotivationCases for which V100 < 85% and D90 < 130 Gy or > 180 Gy are classifiedas being dosimetrically ?suboptimal?, and may be subject to review by aQA committee [35], along with cases of high toxicity to organs at risk, suchas the rectum and urethra. Suboptimal cases are usually re-evaluated onDay-30 with additional CT, and often magnetic resonance (MR), imaging.At some centers in the BCCA Prostate Brachytherapy Program, MRimaging is regularly obtained on Day-30 and incorporated into dosimetryprocedures, for its superior soft tissue edge detecting abilities [9]. This isachieved through the use of coregistration or ?fusion? with CT. HoweverMR is not routinely available at all centers, due to limited resources. Ifobtained, a dose calculation is performed on anatomical structures, manuallycontoured on MR by the RO, based on seeds as located on CT. The twoimage sets are fused by matching CT seeds to the negative seed ?voids? onMR.1.4 Thesis MotivationOne of the significant challenges to accurate dose reporting for postim-plant QA stems from the limitations of using CT for visualization of patientanatomy. The poor soft-tissue contrast observed on CT presents difficultiesin identification of the boundaries of structures such as the prostate, andcan lead to large interobserver and intraobserver variability; see Figure 1.7.This is a well documented problem. The AAPM?s TG-137 report ?III [57]contains a summary and literature review on the choice of imaging modality(ultrasound, CT and MR) and its impact on contoured volume variabilityleading to inconsistencies in reported dose. From early on, concerns overCT-based target delineation and its effect on dosimetric parameters havebeen raised, particularly as compared to the relative consistency found onMR images [6, 24]. In general, prostate volumes defined on CT are largerthan they appear on ultrasound and MR [54, 66, 73], both of which allowsuperior anatomical visualization. Difficulty is most pronounced at the su-perior and inferior aspects: superiorly, confusion stems from the overlapbetween the prostate base and the bladder neck, as well as indistinguisha-bility from the adjacent seminal vesicles, while at the apex, the levator animuscles, genitourinary diaphragm and neurovascular bundles are often mis-takenly included [46, 64, 66]. Furthermore, imaging artifacts from the seedsthemselves, and user bias and subjectivity on the part of the person contour-ing can have a non-negligible effect on organ definition. This has a directimpact on the reliability of dose metrics assessed from CT contours alone121.4. Thesis MotivationFigure 1.7: An example demonstrating the large interobserver variabilitythat can exist in CT contouring. Shown here are an axial CT scan slicelocated towards the prostate base without (top left) and with (top right)contours overlaid. Contours were manually defined by 5 experienced physi-cians. We also show the same contours as seen on a sagittal (bottom left)and coronal (bottom right) representation.[29, 37]. Despite this, CT-based postimplant dosimetry is still currently thestandard of care in prostate brachytherapy.More recently, techniques that include prostate delineation on MR havebeen encouraged by professional organizations such as the AAPM [57] andABS [21], and put into practice at a number of treatment facilities, includ-131.4. Thesis Motivationing some clinics in the BCCA as mentioned above. Typically, this is donethrough fusion of Day-30 MR to CT, with registration performed by eitherseed-to-seed void matching, or comparison of bony and soft tissue anatomy,all leading to greater reproduciblility [3, 19, 42, 62, 85]. However, as alreadyalluded to, the cost and availability of MR prohibits its regular use in allclinics. This motivates our search for a technique that could reduce Day-0 CT-based dosimetric variability, which relies only on resources alreadyavailable in the clinic.TRUS imaging is routinely used during preimplant and intraoperativeprocedures, and provides excellent soft tissue contrast compared to CT (seeFigure 1.8); however, its use in postoperative dosimetric evaluation is lim-ited by poor seed visualization, making it extremely difficult to locate all ofthe seeds following a full implant [30]. Seed segmentation in ultrasound ishindered by specularity (high acoustic reflectivity from the smooth, small-diameter surface), clutter and confusion from other nearby highly reflect-ing objects such as calcifications, and shadowing. It is a well establishedproblem, having been studied and reported on by many groups attempt-ing to devise automated techniques, particularly because of the potentialbenefits for intraoperative dosimetry if coregistered with fluoroscopy [eg.23, 48, 83, 86, 88].Attempts have also been made to fuse ultrasound to CT. Several au-thors have reported on postimplant dosimetry based on CT seed locationsand preimplant TRUS contours, that had been registered using soft tissuelandmarks such as the urethra (with Foley catheterization) or rectal surface[8, 44, 55]. However, since the data are obtained separated by weeks orlonger, during which the patient has undergone a brachytherapy implant,which is known to cause edema [71, 90], significant differences in the patientanatomy are expected to exist between the two. Furthermore, the presenceof the TRUS imaging probe, causing deformation of the patient?s anatomyby compressing the posterior aspect of the prostate, and the fact that TRUSis obtained with the patient in lithotomy position rather than supine, areproblems that persist even if the differences due to edema are reduced byobtaining CT and TRUS closer in time [26]. Dosimetry performed on si-multaneous TRUS and CT images obtained in the same pose, and fusedbased on the transducer itself and refined using seed positions, has beenshown to be feasible [80]; however, it is unclear whether dose metrics in thisset up truly represent the actual dose received by the patient under normal(ie. uncompressed) conditions, not to mention the added inconvenience anddiscomfort experienced by the patient due to additional TRUS imaging.Thus, our objective is to develop a method for improving CT-based141.4. Thesis MotivationFigure 1.8: Midgland prostate slices as seen on intraoperative TRUS (left)and postoperative CT (right). Images are approximately scaled to the samedimensions. The intraoperative TRUS image was obtained after 3 strandshad been implanted, while the postoperative CT was obtained after 22.Note how soft tissue boundaries are distinct on TRUS, but seeds locationsare obscured, while the opposite is true on CT.postimplant dosimetry that uses resources already in place, namely intraop-erative TRUS imaging. Anatomical information based on TRUS will informa mathematical model, which can be registered to CT image coordinates us-ing matched seed positions, and deformed to generate prostate contours thatreflect patient anatomy as seen in CT space. Contours based on a mathemat-ical model, unlike those derived from raw TRUS images, take into accounteffects such as probe compression and edema. Modeling and registration ismade possible by software already part of clinical practice at BCCA [15, 40].TRUS contouring will be aided by sagittal ultrasound, while seed localiza-tion is assisted by intraoperative fluoroscopic imaging, and made feasibleby capturing only a small number of seeds (?20) early on during the im-plant procedure. Except for the additional time required to collect data inthe operating room (typically ?10?15 minutes), patients undergo standardbrachytherapy treatment as they would under normal conditions. We aimto produce consistent contours that can be generated semi-automatically,from which dose metrics can be more objectively derived.15Chapter 2Methods and Materials2.1 Data2.1.1 CollectionStudy subjectsThe data for this research was collected from a cohort of 21 patients un-dergoing prostate brachytherapy at the BCCA Vancouver Center betweenFebruary and September, 2011. Ethics board approval for our study wasreceived, and informed patient consent was obtained from all study sub-jects. Treatment was performed following routine clinical procedures, withstandard materials and equipment. One subject?s data was omitted fromthe analysis for reasons described in ?3, resulting in only 20 case subjects.Radiation SourcesThe main brachytherapy implant sources used were RAPIDStrand (On-cura, Arlington Heights, IL), containing OncoSeed Model 6711 sources. Thesource (or seed) design is a cylindrical titanium capsule 4.5 mm long and0.8 mm in diameter, containing radioactive 125I adsorbed onto a silver rod[59, 65]. The strand and spacing material typically takes 60 to 90 days tobe absorbed. Strands are loaded into 18-gauge hollow needles (1.270 mmouter diameter) for implantation.Recently, a thinner seed and strand model, also by Oncura (ThinStrandModel 9011) [68], was introduced into clinical practice at BCCA. Thin-Strands are planned and used similarly to RAPIDStrand, but are loadedinto 20-gauge needles (0.9081 mm outer diameter).Imaging UnitsUltrasound data were obtained with the BK (BK Medical, Herlev, Denmark)Flex Focus 1202 ultrasound scanner, and transducer Type 8848 containinga convex array for imaging in the transverse plane, and a linear array for162.1. Datathe sagittal plane. The data were acquired in B-mode, which produces 2-Dbrightness pixel maps.Fluoroscopy images were obtained with the GE OEC Series 9800 Plus(GE Healthcare, Little Chalfont, UK) mobile C-arm unit, which producedprojection images of the patient in the coronal plane. CT images wereobtained with the GE Lightspeed RT 16 CT Simulator.2.1.2 ProcedureThe following intraoperative data collection steps were taken during eachsubject?s implant operation. After 3?4 strands had been implanted, corre-sponding to the 14?28 seeds that were to be used as registration seeds, weobtained:? a TRUS volume study,? a sagittal ultrasound scan, on or close to the prostate midline,? five fluoroscopic coronal projections, with the C-arm rotated laterallyin the transverse plane at the following angles: ?10?, ?5?, 0?, 5?and 10? relative to the anterior-posterior axis. The precise angle wasrecorded using a digital protractor.Following the implant, each patient received a pelvic CT scan, as part ofroutine QA procedures, consisting of ?40?60 transverse images separatedby 0.25 cm. All imaging data was either collected as or converted to theDICOM (Digital Imaging and Communications in Medicine) file format.2.1.3 SoftwareThroughout this research, the commercial software VariSeed (Varian Medi-cal Systems, Palo Alto, CA), was used extensively. It is the standard soft-ware used clinically at BCCA for permanent implant prostate brachytherapytreatment planning and dose calculation. VariSeed has many functions, afew of which were particularly important to this research:? manual segmentation of anatomical structures, or ?contouring?, andcontour manipulation? seed localization in a 3-D coordinate system? calculation of isodoses and dose volume histograms for segmentedstructures172.2. InitializationOther software tools developed in-house and used in clinical practice atBCCA were used in this research as well, and will be referred to by the namesAutoContour and PlanReconstruction. AutoContour is a semi-automaticcontouring program that generates prostate contours on the preimplantTRUS volume study based on user-selected initialization points, edge de-tection algorithms, and a 3-D ellipsoidal mathematical model [4, 40]. Plan-Reconstruction performs an automatic 3-D reconstruction of stranded seedtrains segmented on CT images, matching each seed to the correspondingpreimplant planned seed and needle number [15].Software developed by collaborators in the Department of Electrical andComputer Engineering, UBC, was used to reconstruct seed locations in 3-Dcoordinates from coronal fluorscopy images using a back projection algo-rithm (described in ?2.4 of [39]).All other software for this analysis was written in Matlab (Mathworks,Natick, MA).2.2 InitializationHere we describe the initialization steps undertaken to prepare the raw datafor input to the modeling routine.2.2.1 Prostate segmentation in ultrasoundAn initial set of prostate contours on the intraoperative TRUS volume studywas generated following standard clinical guidelines using the program Au-toContour. Initialization points, demarcating TRUS probe center and sixdefined points along the prostate boundary (see [40] for a description), werealso exported. Automatic contours were reviewed and manually modified inVariSeed by an RO according to BCCA treatment planning protocols.A sagittal image of the prostate, obtained to aid in the localization ofprostate boundaries in the base and apex region, was manually contouredby the same RO who reviewed the TRUS contours. We then estimated (byeye, because an exact alignment was not given by the ultrasound imagingunit) a coordinate shift between the sagittal and axial images to align theiranterior-posterior and superior-inferior boundaries as best as possible. TheRO-modified axial contours were then deformed in the anterior-posteriordirection to match the sagittal image, on which the anterior boundary wasmore distinct, especially at the base and apex. No lateral modificationswere made. The adjusted prostate contours formed the basis of the prostatemodel shape.182.2. InitializationFigure 2.1: Sagittal ultrasound images, without (left) and with (right) anRO-defined contour overlaid.2.2.2 Seed segmentation in fluoroscopyIn order to assist with and validate seed segmentation in ultrasound, intraop-erative fluoroscopy images were obtained at the same time as the ultrasoundvolume study. Seeds on fluoroscopic images were localized in 3-D space fromfive coronal projections obtained at known angles of rotation. Note that flu-oroscopy images were obtained with the ultrasound TRUS probe in place,to better match the patient?s anatomical conditions during ultrasound.A manual reconstruction was performed by hand: the seed cloud wasexamined and each seed was assigned the appropriate strand number basedon the preplan. This process was possible because only a small number ofseeds had been implanted at this stage, and their relative positioning wasknown from the treatment plan. The absolute distance scale of the seedcoordinates was not obtained at the time of the implant; however, since wewere primarily interested in their position relative to ultrasound, this wasnot necessary. To assign an absolute distance scale to seed coordinates, weused the fact that neighbouring seeds on a strand had known separations.2.2.3 Seed segmentation in ultrasoundIntraoperative seeds were manually segmented on TRUS, guided by the fluo-roscopic seed reconstruction, the expected preplan location of seeds relativeto the implantation template, and the known seed separation. VariSeedwas used to facilitate this process: each slice was examined, and seeds were192.2. InitializationFigure 2.2: A fluoroscopy coronal projection, containing 22 seeds. TheTRUS probe and Foley catheter are also visible.assigned a strand number and coordinate located on the slice that they ap-peared the brightest on. Occasionally, the end of a strand extended beyondthe first or last slice of the volume study. Most often, this occurred towardsthe inferior, due to dragging by the needle as it was retracted from thepatient, although there are other obstacles that prevent implants from per-fectly reproducing the treatment plan in practice (eg. see ?1.2.5 of [16] foran overview). In the ?50% of cases for which this occurred, only a subsetof seeds could be segmented. Omitted seeds (typically 2?3) were identi-fied, using the fluoroscopy reconstruction as a reference when available, andtaken into account in subsequent seed matching procedures. Intraopera-tive ultrasound seed locations were exported, and used in the modeling forregistration purposes.202.3. Model AlgorithmValidationDue to the large expected uncertainties in seed segmentation on TRUS im-ages, we validated seed reconstruction by comparison to fluoroscopy. Seedsthat could be identified on both modalities were matched, and a registrationbetween their 3-D coordinates was computed, using singular value decom-position [13]. The measured rotation and translation was applied to thefluoroscopy seed coordinates to produce the best possible alignment withcorresponding ultrasound seeds, and a mean 3-D offset value was calculatedfrom all known pairs. This represented the estimated error on TRUS seedpositions.2.2.4 Seed segmentation in CTPostoperative CT images were obtained as part of routine QA procedures.A 3-D seed reconstruction of the completed implant was generated basedon the preimplant treatment plan using the in-house clinical software Plan-Reconstruction [15]. In the clinic, physicists ensure that a complete planreconstruction is achieved, ie. that all seeds are accounted for. Having thisallowed us to identify precisely which, out of the 80?150 seeds in the full im-plant, belonged to the 3?4 strands containing our 14?28 registration seeds.Reconstructed CT seed locations were exported and used in the modelingfor registration.2.3 Model AlgorithmIn this section, the steps used to create an ultrasound-based model of theprostate are described. A flowchart summarizing the main components isprovided in Figure 2.3. We generated model contours via two differentstreams, which from here on will be referred to as model X and Y. Manyof the steps are the same for both versions, and where they diverge is illus-trated in Figure 2.3. We indicate in the titles of the following subsectionsif the procedure was performed for one model set only; otherwise, they arecommon to both.The procedures to generate a 3-D prostate model were devised by S.Badiei [4] and S. Mahdavi [40], and the procedure to match and register andseed coordinates was written by N. Chng [15]. Details of these algorithmshave been described extensively in their publications. What follows is a briefsummary of the main components of our procedure, and we have highlightedin Figure 2.3 the components unique to this study.212.3. Model AlgorithmFigure 2.3: Flowchart illustrating the two modeling streams: the taperedellipsoid model X, and linear interpolation model Y. We have emphasizedin bold the modeling steps that are unique to this project.The input data, having been prepared as described in ?2.2, include DI-COM images, contours, seed locations and identities, initialization parame-ters, and the ultrasound sagittal-axial coordinate shift.In model coordinates, the x-axis refers to the lateral direction, increasingfrom patient right to left, the y-axis refers to the anterior-posterior direction,increasing towards the posterior, and the z-axis refers to the superior-inferiordirection, increasing towards the superior.222.3. Model Algorithm2.3.1 UnwarpingThe presence of the TRUS probe causes a posterior deformation in patientanatomy, that we refer to as ?warping?. Contour and seed coordinates onultrasound images must be unwarped, since the probe effect is not presentin the CT image, and because the modeling algorithm requires that a 2-Delliptical fit be performed to the contour points. The probe is assumed tocause a radial compression that is maximal at the posterior medial aspectof the prostate, and decreases in amplitude with distance from the probecenter, and with angle away from the midline. In polar coordinates with theorigin at the probe center, the sinusoidal Gaussian functionrnew = r ? r sin(?) exp(?r22?2)(2.1)describes the radius rnew of an unwarped point originally with radius r,where ? is the angle from the horizontal axis (? = 90? along the midline).? is the stretch variable in the radial direction, and is measured from theinput initialization parameters [4]. Assuming a uniform deformation alongthe probe axis, equation 2.1 and measured ? can be used to unwarp all inputcontour points and seed locations.2.3.2 Model X: Untapering and ellipse fittingOnce the warp compression has been removed, prostate anatomy as seen ontransverse ultrasound resembles an ellipse that often tapers towards the an-terior. This 2-D contour can be described as an ellipse containing a taperingparameter t1, and represented as(x?ax)2+(y?ay)2= 1 (2.2)x? =(x? x0)t1ay(y ? y0) + 1y? = (y ? y0)where x0, y0 are the ellipse center coordinates, and ax, ay are the radii onthe x and y axes, respectively [40]. If t1 = 1, the anterior tip comes toa point at a 90? angle; if t1 = 0, the equation describes a straightforwardellipse (see Figure 3 in [40] for an illustration). Equation 2.2 was fit to eachunwarped axial contour to find t1; this is similar to the original procedure232.3. Model Algorithmin AutoContour, except here the fit is performed on intraoperative contoursrather than on edge detection boundaries. Contours were then untapered,and a 2-D ellipse was fit to the unwarped, untapered contour points on eachtransverse ultrasound slice.2.3.3 Model X: 3-D tapered ellipsoid fittingIn order to create a smooth, continuous 3-D surface, an ellipsoid that tapersin the ?z direction (ie. towards the apex) is fit to the 2-D ellipses. Thecross-section of the tapered ellipsoid is itself an ellipse. The tapered ellipsoidis defined as (x?ax)2+(y?ay)2+(z?az)2= 1 (2.3)x? =(x? x0)t2az(z ? z0) + 1y? =(y ? y0)t3az(z ? z0) + 1z? = (z ? z0)where x0, y0, z0 are the coordinates of the ellipsoid center, ax, ay, az are theradii on the x, y and z axes, respectively, and t2, t3 are the respective taperingparameters in the x and y directions along the axis of the probe [40].2.3.4 Model X: Slicing and taperingHaving found the tapered ellipsoid model parameters, we can construct afinely-spaced grid of points representing the prostate surface. First, thecenters of the 2-D ellipses from ?2.3.2 are used to define a line in the 3-Dcoordinate system. This line will define the centers of the final contoursdrawn from the tapered ellipsoid model, accounting for any possible pitchand yaw that may exist in the main axis of the ultrasound prostate anatomyrelative to CT anatomy.Next, the tapered ellipsoid model is finely sliced in the z-direction, gen-erating a regular grid of closely spaced ellipse boundary points at small sliceseparations. In order to construct a model surface that closely resembles theultrasound anatomy as much as possible, we re-apply the anterior taperingthat was removed in ?2.3.2. The parameter t1 at the z-slice locations is es-timated by interpolating between the values measured on TRUS slices, andused to apply tapering to model ellipses in the anterior direction. We defineour model surface as this fine grid of points.242.3. Model Algorithm2.3.5 Model Y: Linear interpolationA second ultrasound-based model was generated without fitting the surfaceto a tapered ellipsoid, but by using simple linear interpolation. This model,referred to as ?model Y?, was primarily motivated by oncologists? interest inseeing how spatially transformed TRUS contours would appear on CT im-ages with minimal manipulation applied. Note that the posterior unwarpingdeformation is treated in the same way for this model.Without the use of 3-D ellipsoid model parameters, model Y can not bedescribed on a fine spatial scale, being limited by sparsely sampled TRUSslice contours. The midline sagittal contour, however, can provide impor-tant information about the prostate base and apex regions missing fromthe transverse volume study. We define the superior-most and inferior-mostpoints of the sagittal contour as the base and apex, respectively, and acoarse surface model is constructed by linearly interpolating between thebase, unwarped axial contours, and apex.2.3.6 Registration and edemaTo transform the prostate model from ultrasound to CT coordinates, a seed-based rigid registration is calculated between the unwarped TRUS and CTseed positions. Since the seed correspondence is already known, comput-ing the rotation and translation factors is relatively straightforward usingsingular value decomposition [13]. The computed transformation is then ap-plied to the ultrasound model and seeds: thus, the model surface and seedpositions are defined in CT image coordinates.Expansion of the prostate due to edema is expected to occur betweenintraoperative ultrasound and postoperative CT. We estimate the edemafactor as the ratio of the mean radial components of matched seeds arounda point defined as the center of expansion (COE): this assumes a uniformexpansion around COE. For model X, the COE is defined as the best-fitcenter of the 3-D tapered ellipsoid x0, y0, z0 from ?2.3.3; for model Y, theCOE is defined as the geometric center of the coarse surface model from?2.3.5. If the edema factor is >1, radial expansion is applied to modelcoordinates.2.3.7 CT contour slicingFinal CT contours are generated by extracting axial slices of the transformedmodel at the z coordinates corresponding to the axial CT scan slices. This252.4. Evaluationproduces a set of x, y contour coordinates that can be imported and overlaidon CT images in VariSeed or other software for further analysis.2.4 EvaluationRoutine clinical procedures for prostate brachytherapy patients require thetreating RO to contour patient anatomy on postoperative CT for dose cal-culation purposes. We collected these prostate contours, and also obtained 4additional sets of CT prostate contours for each patient from volunteer physi-cians who were recruited to participate in this research. All volunteers wereexperienced physicians trained in prostate segmentation, and were blindedto each others? contributions, as well as to other imaging modalities (ie.ultrasound) and patient identity.Thus, our data set for analysis consists of prostate contours for 20 pa-tients (or subjects), from 7 ?observers?, which includes both manual andmodel contour-generators.2.4.1 STAPLEThe Simultaneous Truth and Performance Level Estimation (STAPLE) al-gorithm [87] uses an iterative expectation-maximization algorithm to simul-taneously compute a probabilistic estimate of the true segmentation, and aquantitative assessment of the performance level of each ?rater?, based onthe input from multiple raters. The true segmentation is an unknown binaryvariable Ti ? {0, 1} for each voxel i of N total voxels in the CT scan volume,and is either present (1) or absent (0). For each subject, we input to thealgorithm a collection of voxel sets, each representing the CT image prostatesegmentation ?decisions? by one of the 5 manual contour generators. Deci-sions are also binary Di ? {0, 1}, meaning that voxels are either included inthe manually contoured region or not, and are directly observable.At each iterative step, STAPLE evaluates a weight variable Wi, repre-senting the conditional probability of the true segmentation at each voxel ibeing equal to 1, given the decisions Di and an estimate of the rater perfor-mance from a previous iteration. Wi is a continuous parameter between 0and 1, ie. Wi ? [0, 1], and is defined asW (k?1)i ? f(Ti = 1|Di, p(k?1), q(k?1)) (2.4)where k is the iteration number, and (pk, qk) is the estimate of the perfor-mance level parameters at iteration k. Rater performance parameters are262.4. Evaluationalso evaluated at each iteration, based on the previous iteration?s estimatedtrue segmentation. The sensitivity is defined as pj = Pr(Dij = 1|Ti = 1)and the specificity as qj = Pr(Dij = 0|Ti = 0), where pj , qj ? [0, 1]; thesecharacterize the quality of each rater j = 1, . . . , 5. Iteration ends when Wiconverges.Following the final iterative step, we obtain a consensus segmentation bychoosing a threshold value of Wi. To assess the quality of rater performance,the sensitivity and specificity parameters pJ , qJ based on the final Wi areevaluated for all observers, of which there are now J = 1, . . . , 7 in total(5 manual plus 2 model). Additional performance parameters evaluated arethe statistical measures known as the predictive values, defined as PVJ(s) =Pr(Ti = s|DiJ = s), ?s ? 0, 1. The positive predictive value corresponds tothe case s = 1, the negative predictive value to s = 0. From Bayes theorem,we can express PVJ(s) in terms of the known quantities W , p and q:PVJ(1) = Pr(Ti = 1|DiJ = 1)=Pr(Ti = 1) ? Pr(DiJ = 1|Ti = 1)Pr(Ti = 1) ? Pr(DiJ = 1|Ti = 1) + Pr(Ti = 0) ? Pr(DiJ = 1|Ti = 0)=f ? pf ? p+ (1? f) ? (1? q)andPVJ(0) = Pr(Ti = 0|DiJ = 0)=Pr(Ti = 0) ? Pr(DiJ = 0|Ti = 0)Pr(Ti = 0) ? Pr(DiJ = 0|Ti = 0) + Pr(Ti = 1) ? Pr(DiJ = 0|Ti = 1)=(1? f) ? q(1? f) ? q + f ? (1? p)wheref = Pr(Ti = 1) =1NN?iWiis the global probability of the true segmentation, and W , p and q areunderstood to be binary random variables, ie. the conditional probabilityf(Ti = 0|Di, p, q) is 1? f(Ti = 1|Di, p, q).The relationship between true segmentation voxels, rater decisions, andperformance parameters is summarized in Table 2.1. Later in this work, theterminology True Positive (TP), False Positive (FP), True Negative (TN),and False Negative (FN) will be used to discuss those fractions of voxelsthat relate to the STAPLE parameters.272.4. EvaluationT = 1 T = 0D = 1 TP FPPV(1) =Pr(T = 1|D = 1)D = 0 FN TNPV(0) =Pr(T = 0|D = 0)Sensitivity Specificityp = Pr(D = 1|T = 1) q = Pr(D = 0|T = 0)Table 2.1: Summary of statistical terminology and the general relationshipbetween parameters.Statistical comparisonOur goal is to determine whether model-generated CT contours of theprostate can be considered equivalent in quality to those manually-generated,knowing that there is often considerable variability in manual contouring.The benefit of possessing multiple manual contours for each subject is thatthe interobserver variability can therefore be quantitatively measured. Foreach subject and STAPLE performance parameter, we estimate the 95%confidence interval (CI) around the sample mean value of the manual seg-mentations, from the estimated standard error of the mean s?x? and the criti-cal value from the Student?s t-distribution t? corresponding to the ? = 0.05level of significance for a non-directional (two-tailed) test:95% CI = x?? t?s?x?. (2.5)We use the t-distribution in determining the range, which makes the simpleassumption that within-subject interobserver measurements are normallydistributed. The performance parameters evaluated for model X and Y arethen compared against this confidence interval; if within the 95% CI, themodel is classified as being ?in agreement?.2.4.2 Volumetric and dosimetric analysisDosimetric parameters are ultimately what determines the success or fail-ure of a particular treatment in terms of QA. Given each observer?s set ofcontours, the total prostate volume and corresponding V100 and D90 can besimply computed using the treatment planning software VariSeed. Within-subject interobserver 95% CIs for manual segmentations are evaluated usingthe same method as ?2.4.1 for the volumetric and dosimetric parameters (ie.volume, V100 and D90) as a comparison standard against which to comparethe model-generated parameters.282.4. EvaluationSagittal viewAxial viewBase Midgland ApexAnteriorPosteriorLateralLateralFigure 2.4: Schematic diagram illustrating the 9 sector divisions.Analysis of the dose delivered to prostate volume as a whole is of limitedusefulness because no information about the spatial distribution of doseis given. Therefore, we performed a region-specific analysis by dividing theprostate gland into 9 sectors, similar to the analysis of Mahdavi et al [40, 41].First the base, midgland and apex regions are defined, and then each isfurther subdivided into the anterior, lateral (containing both right and leftquadrants together) and posterior regions; see Figure 2.4. The referencevolume chosen to define the sector axes was the STAPLE consensus volumefrom ?2.4.1, since it represents the best estimate of the ?true segmentation?.Thus, the same set of axes in absolute CT coordinates are used to createsectors for all observer segmentations, allowing us to compare ?apples toapples?. Similar to the total prostate analysis, we compare model volume,V100 and D90 parameters to the 95% CI of the manual segmentation meansfor each sector.The 9 sectors are named the BAS, MAS, AAS, BLS, MLS, ALS, BPS,MPS and APS, where the first letter in this convention stands for base,midgland, or apex, and the second letter stands for anterior, lateral, orposterior. When viewed sagittally, the sector divisions appear as they do inFigure 2.5, which we include for ease of reference.It should be emphasized that while differing degrees of variability areexpected in each sector, we are not interested in absolutely quantifying orcomparing them. Rather, we wish to determine whether model-generatedcontours of the whole prostate and its sectors are generally within the ob-served variability.292.4. EvaluationSagittal viewML6 AL6BL6MP6MA6BA6 AA6BP6 AP6Figure 2.5: Schematic showing the 9 sectors labeled as visualized on a sagit-tal view of the prostate.2.4.3 Statistical testingWe also perform traditional statistical tests on the total prostate volume anddosimetric parameters, which is the typical route taken for studies of inter-observer variability. However, simple tests, such as Student?s t-test, are notoptimal in data such as ours because in order to test for differences betweenobservers, results from multiple subjects (which naturally have wide rangesof variability) must be combined. Such inter-patient variability could dom-inate over the interobserver variability we are interested in. Furthermore,it is probable that the values of some parameters of interest, such as V100,are not Normally distributed across subjects, so use of parametric tests thatassume a Gaussian data set is inappropriate.We therefore explore the Friedman test, which is a repeated-measuresmultivariate test for non-parametric data that utilizes ranks; Bewick et al.[5] demonstrate a worked example. In this test, one can detect differencesbetween k different experiments, sometimes called ?treatments?, where eachexperiment is performed once on b different sample groups or ?blocks? ofdata. In our study, patients are blocks, observer contours are treatments,and the volume, V100 or D90 are the values obtained from each treatment302.4. EvaluationBlocks Treatments (Contours)(Patients) 1 2 ? ? ? k1 X11 X12 ? ? ? X1k2 X21 X22 ? ? ? X2k...............b Xb1 Xb2 ? ? ? XbkTable 2.2: Arrangement of raw data for the Friedman Test.performed on each block. The actual numerical values measured are unim-portant, because data are ranked between 1 and k within each block. Anotable feature of the Friedman test is that it assumes differences betweenblocks are large, so no comparisons between blocks are made.In this analysis, the data take the form of b ? k mutually independentrandom variables, with Xij being the observation in block i = 1, 2, . . . , bassociated with treatment j = 1, 2, . . . , k. Table 2.2 demonstrates how dataare arranged.Values within each row i (block) are assigned a rank, denoted by R(Xij),from 1 to k, and the sum of ranks from the jth column (treatment) isRj =b?i=1R(Xij). (2.6)We wish to test the null hypothesis H0: that the order of rankings withina block are equally likely, ie. each treatment has equivalent effect. Thealternative hypothesis H1 is that at least one treatment yields values thatare consistently higher or lower than one other. The Friedman test statisticis defined asT =12bk(k + 1)k?j=1R2j ? 3b(k + 1). (2.7)For large sample sizes, as in the case here, we reject H0 at the ? level ofsignificance if T is greater than the critical value, approximated by the ?2distribution for k ? 1 degrees of freedom.Post-hoc analysisIf the null hypothesis is rejected, we then perform a post-hoc multiple com-parisons test, which identifies which pairs of treatments differ significantly.The procedure requires calculating pairwise absolute differences of the rank312.4. Evaluationsums given by equation 2.6. All pairs of treatments u, v simultaneously dif-fer at the experiment-wise error rate ? for which the following inequality istrue:|Ru ?Rv| ? r(?, k, b), (2.8)where r(?, k, b) is the critical value defined by the particular test beingperformed; we examine results from two tests, described below. The to-tal number of possible pairwise comparisons equals k(k ? 1)/2, where u, vcombinations are given by u < v, u = 1, . . . , k ? 1, v = u+ 1, . . . , k.The Nemenyi test, described in detail by Hollander et al. [31], is consid-ered conservative, ie. less able to detect small differences. Assuming largesample statistics, the critical value is given asr(?, k, b) ? q(?, k,?)?bk(k + 1)12(2.9)where q(?, k,?) is the studentized range statistic (Table A.10, [31]) associ-ated with ? degrees of freedom within groups (ie. large b).The Fisher least significant difference (LSD) test modified for nonpara-metric data, described by Conover [18] and used by Bewick et al. [5], is lessconservative, and defines the critical value asr(?, k, b) = t????????2??bk?j=1b?i=1[R(Xij)]2 ?k?j=1R2j??(b? 1)(k ? 1)(2.10)where t? is the 1 ? ?/2 quantile of the Student?s t-distribution with (b ?1)(k ? 1) degrees of freedom (Table A.21, [18])32Chapter 3ResultsProstate models based on two ultrasound-informed methods were success-fully generated for 20 case subjects. Sagittal delineation was not possiblein one of the original 21 cases due to poor image quality, so this data setwas omitted from the rest of the analysis, as mentioned in ?2.1.1. Registra-tion from TRUS to CT coordinates was performed based on matched seedlocations, and the model was sliced at 0.25 cm intervals corresponding tothe CT scan spacing. Shown in Figure 3.1 are example models generatedusing the 3-D tapered ellipsoid (model X) and linear interpolation (modelY) methods.Figure 3.1: Model X (left) and Y (right) final products are shown as coloursurfaces, with the CT contours overlaid in coloured lines. Matched CT (solidsquares) and ultrasound seeds (open squares) are indicated by a solid lineconnecting them. Note that extraprostatic seeds are intentionally placed inthe posterior to achieve the treatment planning goals of delivering a highdose in the posterior peripheral zone.In the following evaluation, patients will be referred to by a number from1 to 20, manual CT contours are referred to by a letter from A to E, and333.1. STAPLEthe two model contours by X and Y.Ultrasound seed validationSeed localization in ultrasound was aided by fluoroscopy whenever possible.Table 3.1 lists the total number of seeds Ntotal implanted for each patientat the time of intraoperative data collection, and the number that couldbe found and identified on the TRUS volume study NTRUS and coronalfluoroscopy projections Nfluoro. An average of 93% (range 80%?100%) ofseeds were successfully reconstructed on TRUS. In 4 out of the 20 cases, thefluoroscopy seed cloud could not be reconstructed at all; this may have beendue to the use of ThinStrands which were noticeably more difficult to detect,or a misalignment of the C-arm out of the patient?s transverse axis causingproblems with the simple back projection algorithm. In the remaining 16cases, an average of 85% (range 50%?100%) of seeds were reconstructed,with failures possibly resulting from the obscuring effect of the ultrasoundprobe?s presence in the coronal projection.Matched seed coordinates in TRUS and fluoroscopy were registered andthe 3-D relative mean offset was calculated. On average, the number ofseeds for which a TRUS-fluoroscopy correspondence was made Ncorr is 80%(range 50%?100%). Using the fluoroscopy positions as the reference frame,we estimate the error in ultrasound seed locations from the average meanoffset, which was found to be 0.24 cm (range 0.07?0.49 cm).3.1 STAPLEFrom the STAPLE weight parameter Wi, we estimate the true segmentationfrom the consensus volume defined at the cutoff threshold of Wi ? 0.5,the value used in an example by Warfield et al. [87]. Although seeminglyarbitrary, it was observed that the evaluated voxel weights were generallybinary in nature, tending towards values either >0.9 or <0.1; therefore, theconsensus volume is highly insensitive to the chosen threshold of Wi (seeFigure 3.2).Having assigned a value of Ti = 0 or 1 to each CT voxel, the consensusvolume representing the STAPLE ?truth? is established. From this, per-formance parameters are calculated for all 7 observers from the definitionsgiven in ?2.4.1. It was found that because of the inherent large true negative(TN) fraction by nature of the CT voxel space being so large, the specificity,and therefore the negative predictive value, were always near unity; that is,qJ = Pr(DiJ = 0|Ti = 0) ? 1 and PVJ(0) ? 1 for all J . Thus, we focus on343.1. STAPLEPatient Thin Ntotal NTRUS Nfluoro Mean offset (cm) Ncorr1 14 14 14 0.10 142 15 12 12 0.07 123 Y 20 17 18 0.28 154 22 20 22 0.14 205 20 16 13 0.12 116 20 18 20 0.13 187 20 19 - - -8 22 18 22 0.20 189 Y 20 20 10 0.12 1010 22 22 19 0.25 1911 28 28 28 0.39 2812 22 20 22 0.28 2013 24 24 14 0.44 1414 Y 22 22 - - -15 Y 20 20 - - -16 22 20 - - -17 Y 20 19 15 0.31 1418 Y 28 25 17 0.49 1419 20 20 19 0.2 1920 18 18 18 0.24 18Table 3.1: Results of TRUS and fluoroscopy seed cloud reconstruction andcomparison. Also indicated under column marked ?Thin? are ThinStrandimplant cases. Ntotal is the total number of seeds known to be present duringintraoperative data collection, and NTRUS and Nfluoro are the number ofTRUS and fluoroscopy seeds successfully reconstructed, respectively. Themean offset between registered, matched seeds is calculated for Ncorr numberof seeds for which a correspondence was found.the sensitivity pJ and positive predictive value PVJ(1). To better under-stand the distinction between them, consider the following definitions: thesensitivity is the probability that the observer chose all the true structurepresent, while the positive predictive value is the probability that structureis truly present when the observer chose it. Sensitivity is synonymous with?recall? and PV(1) with ?precision? in information retrieval terminology.The results of STAPLE parameter analysis for all 20 subjects are plottedin Figure 3.3, and summarized in Table 3.2. The sensitivity parametersmeasured for both model-generated segmentations are in agreement with353.1. STAPLEFigure 3.2: An example set of evaluated STAPLE weight parameter Wimaps. Note that voxel values tend to be nearly 0 or 1, so the consensusvolume is largely independent of our choice of Wi threshold.the manual in nearly all cases, while the positive predictive value agreesin only 6 (model X) and 8 (model Y) cases, tending towards values lowerthan the manually determined 95% CI. A similar sensitivity with relativelylow PV(1) can be interpreted as follows. Manual and model observers aregenerally contouring (and omitting) the same number of voxels inside theconsensus volume, meaning the TF and FN numbers are similar. However,the model is selecting a larger fraction of voxels outside the consensus region(FP) than the typical manual contourer, ie. the model is ?overcontouring?relative to the consensus segmentation. This is illustrated in Figure 3.4.363.1. STAPLE0.  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 200.70.750.80.850.90.9511.051.1Positive predictive valuePatientFigure 3.3: Results of STAPLE analysis. Mean sensitivity p and positivepredictive values PV(1) and 95% CIs derived from manual contours areshown (blue crosses and error bars), along with model X (red circles) and Y(green squares) values.373.2. Volumetric and Dosimetric AnalysisModel Sensitivity PV(1)X 20 6Y 19 8Table 3.2: Number of cases for which the model STAPLE parameter waswithin the 95% CI variability range of manual values, out of a maximum 20.3.2 Volumetric and Dosimetric Analysis3.2.1 Total prostate dosimetryThe measured volume and dosimetric parameters for the total prostate vol-ume are summarized in Table 3.3, and plotted in Figure 3.5. We includeall values in the Appendix for completeness; see Tables A.1, A.2 andA.3.In 19 cases for model X, and 18 for model Y, out of 20, the prostate vol-ume generated by modeling methods are within the expected variabilityrange determined through manual segmentation, while there is only mod-erate agreement between model and manual dosimetric parameters. Modelcontours tended to produce slightly larger volumes compared to manualsegmentation.Model Volume V100 D90X 19 9 10Y 18 12 13Table 3.3: Number of cases for which the model total volume and doseparameter was within the 95% CI variability of manual values, out of amaximum 20.It is notable that at least one case (Patient 19) designated ?suboptimal?by routine QA classification schemes, ie. V100 < 85% and D90 < 130 Gybased on manual contouring, would have received an acceptable assessment,and subsequently not been subject to review, had the contours been modelgenerated. Conversely, D90 for Patients 2 and 12 yields a borderline dosime-try assessment for at least one of the model-generated contours, whereas theaverage manual contour dose assessment was acceptable.3.2.2 Sector dosimetryWe break down the analysis into 9 sectors. Plots of the mean volumetricand dosimentric parameter value and 95% CIs for each patient and sector383.2. Volumetric and Dosimetric Analysis';'A'%'&7Figure 3.4: Example illustrating high sensitivity with low positive predictivevalue. Shown are contours generated by three manual ?decisions? DA, DB,and DC , the consensus region representing the estimated true segmentationT , and one model decision DX . In this scenario, all decision contours resultin roughly the same sensitivity, but by selecting a proportionally larger areaoutside of T , DX will have a relatively smaller positive predictive value.(similar to Figure 3.5) are presented in the Appendix, see Figures A.1?A.9.On a handful of occasions for both manual and model generated contours,the measured volume for a particular sector was zero: this is a consequence ofthe sector divisions for all prostate segmentations being defined in absolutecoordinates based on the STAPLE prostate axes. Null results were excludedfrom the analysis. All sector results are summarized in Table 3.4. To helpvisualize the results, we present all 9 sectors in a single diagram for eachparameter (volume, V100 and D90) and modeling method (X and Y), usinga colour scale to represent level of agreement, in Figure 3.6. For ease ofreference, refer to Figure 2.5 which illustrates the sector locations as theyappear in the sagittal view.Upon examination, we find that the apical posterior sector (APS) ex-hibits the least agreement between model and manual contours. Model-generated volume and dose parameters for fewer than half of the cases were393.3. Statistical TestingSector Model Volume V100 D90BASX 11 13 11Y 10 12 13MASX 13 17 14Y 10 15 13AASX 16 16 16Y 16 16 16BLSX 15 11 13Y 16 13 13MLSX 14 11 11Y 15 11 12ALSX 15 17 15Y 16 17 14BPSX 18 14 15Y 18 16 15MPSX 14 11 9Y 15 12 10APSX 9 10 9Y 8 9 9Table 3.4: Numbers of cases for which the model sector volume and doseparameter was within the 95% CI variability of manual values, out of amaximum 20.within the manual 95% CI: 9 cases for volume, 10 for V100 and 9 for D90for model X, and 8, 9 and 9 cases for model Y, respectively. Figure A.9 re-veals that in cases of poor agreement, modeling often tends to overestimatethe volume relative to manual contours by including regions further to theposterior and inferior of manually-defined boundaries, see Figure 3.7. Thishas the effect of decreasing the associated V100 and D90 to this sector. To alesser degree, the same model overestimation of the volume is observed forthe midgland posterior sector (MPS) as well (see Figure A.8), although thenumber of cases for which values exceeded the 95% CI variability range isless extreme.3.3 Statistical TestingWe test for statistical differences in volumetric and dosimetric parametersbetween observers. Evaluation up to this point has been focused on mea-403.3. Statistical Testingsuring how frequently model-determined parameters are in agreement withmanual. Statistical testing on the other hand quantifies the significancelevel of their agreement, and importantly, can be used to determine if eithermodel?s segmentation is more significantly different from the group than anyone manual segmentation.We perform the Friedman test and post-hoc analysis outlined in ?2.4.3on the volume, V100 and D90 values. Values and ranks for each subject(blocks, b = 20) produced by each observer?s contours (treatment, k = 7)are presented in the Appendix, see Tables A.1?A.6. Results are reportedin Table 3.5. The Friedman test statistic T is calculated with 6 degrees offreedom and the corresponding two-tailed p-values are shown. For all param-eters, we reject the null hypothesis at a significance level p < 0.01, meaningthat contouring between all observers and models are not equivalent.Friedman Post-hocParameter T p-value Nemenyi?s r Fisher?s LSD rVolume 61.8857 <0.0001 40.2860 19.3244V100 18.2786 0.0056 40.2860 25.5664D90 26.9357 0.0001 40.2860 24.4543Table 3.5: Results of statistical tests.Post-hoc tests reveal which observer?s segmentations differ at the cho-sen level of significance ? = 0.05. For Nemenyi?s test, given by equa-tion 2.9, we obtain from tables the appropriate Studentized range statisticq(0.05, 7,?) = 4.170 for infinite degrees of freedom, and calculate a criticalvalue of r(0.05, 7, 20) = 40.2860 for all three parameters. Using the modifiedFisher?s LSD test, we obtain from tables t0.05 = 1.9810 for 114 degrees offreedom, and calculate the critical values from equation 2.10, reported inTable 3.5.We find the absolute differences in the sum of ranks (equation 2.6) be-tween each pair of observers, for comparison against the minimum criticaldifferences calculated from post-hoc tests. All pairwise differences for vol-ume, V100 and D90 are presented in Tables 3.6, 3.7 and 3.8, respectively.Highlighted are those whose difference in rank sum are greater than Fisher?sLSD (in bold) and Nemenyi?s (with asterisk) critical values, and thus con-sidered significantly different.With Fisher?s LSD test being highly sensitive to small effects, we findthat most of the pairwise differences across our three parameters are consid-ered significant, which makes it difficult to identify any standout observers.The more conservative Nemenyi test identifies a smaller subset of those re-413.3. Statistical Testingsults that are the most significant. Focusing on Nemenyi?s test results, weexamine the volume parameter and find that model X is identified as signif-icantly different in 1 pair, and model Y in 2. This is less than the manualsegmentation C, which significantly disagrees in all 6 possible pairwise com-binations. The V100 parameters for model X and Y are not found to besignificantly different relative to any other observers. For D90 paramter,model Y is significantly different from observer C, but so too are observer Aand E. In fact, the overall trend indicated by post-hoc analysis is that ob-server C is the outlier. Removing C from the comparison, it is still apparentthat X and Y are not more different than any other observer by either teston any of the parameters: observers D (in volume), A and E (in V100 andD90) are equivalent to or exceed X and Y in numbers of significant pairwisedifferences.A B C D E X YA - 1 63* 36 2 20 13B 1 - 64* 35 3 19 14C 63* 64* - 99* 61* 83* 50*D 36 35 99* - 38 16 49*E 2 3 61* 38 - 22 11X 20 19 83* 16 22 - 33Y 13 14 50* 49* 11 33 -Table 3.6: Absolute differences in the rank sums for the whole prostatevolume between all pairs of observers. Note that values are symmetric acrossthe main diagonal. Values in bold are greater than Fisher?s LSD criticalvalue 19.3244 for volume; values marked with an asterisk (*) are greaterthan Nemenyi?s critical value 40.2860.423.3. Statistical TestingA B C D E X YA - 35 44* 38 8 34 16B 35 - 9 3 27 1 19C 44* 9 - 6 36 10 30D 38 3 6 - 30 4 22E 8 27 36 30 - 26 8X 34 1 10 4 26 - 18Y 16 19 30 22 8 18 -Table 3.7: Absolute differences in the rank sums for total prostate V100parameter between all pairs of observers. Refer to Table 3.6 caption for adescription; here, Fischer?s LSD critical value is 25.5664.A B C D E X YA - 40 54* 40 5 31 12B 40 - 14 0 35 9 28C 54* 14 - 14 49* 23 42*D 40 0 14 - 35 9 28E 5 35 49* 35 - 26 7X 31 9 23 9 26 - 19Y 12 28 42* 28 7 19 -Table 3.8: Absolute differences in the rank sums for total prostate D90parameter between all pairs of observers. Refer to Table 3.6 caption for adescription; here, Fischer?s LSD critical value is 24.4543.433.3. Statistical Testing1020304050607080Volume (cc)5060708090100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2080100120140160180D90 (Gy)PatientDose metrics for TotalFigure 3.5: Volume and dosimetric parameter results for total prostate.Mean manual contouring values and 95% CIs (blue crosses and error bars)are shown along with values produced by model X (red circles) and model Y(green squares). Overplotted are the dosimetric ?sub-optimal? QA thresh-olds (dotted lines): V100 = 85% and D90 = 130 Gy and 180 Gy.443.3. Statistical TestingFigure 3.6: Colour sector diagrams representing the level of agreement involume, V100 and D90 for models X and Y. The colour scale on the rightindicates the number of cases, out of a maximum 20, for which the modelparameter was within the 95% CI variability of manual values, in each sectoras seen from the sagittal view (refer to Figure 2.5).453.3. Statistical TestingFigure 3.7: Sagittal, axial and coronal view of a patient?s CT scan, withSTAPLE (cyan) and model X (magenta) contours overlaid. Seed locationsare also marked by cyan bars. The apical posterior sector (APS) tends tobe larger on model-generated contours than it is typically defined manually.46Chapter 4DiscussionThe goal of this study is to determine if an ultrasound-based model of theprostate can provide assistance in performing post-implant dosimetry on CTimages. We have generated model-based contours on CT images and evalu-ated their utility through comparison with manual CT contours, the currentstandard of care at BCCA and the most direct and cost-effective method fordosimetric evaluation as stated in TG-64 [94], despite the known segmenta-tion difficulties (see also TG-137 [58]). Overall, model-based volumes agreewell with manual contours, tending to be slightly larger than the manualmean, which partly motivated our region-specific analysis. Dosimetric pa-rameters V100 and D90 agree in only a moderate number of cases, althougheven when model values are not within estimated manual variability, theydo not fall far beyond 95% CIs. In cases where this is not true (for exam-ple, model contours for Patient 15 show significant deviation from manualvalues), a low STAPLE positive predictive value confirms that the modelsegmentation suffers from poor precision.Comparing the two modeling streams, we observe that STAPLE, volumeand dosimetric parameter results are very similar, although the taperedellipsoid modeling method (X) produces slightly larger volumes than simplelinear interpolation (Y). This is unsurprising, given that the ellipsoid modelfits a smooth, convex surface shape to the known contour points. This resultis also reflected in the higher sensitivity (ie. recall) parameters for model Xfrom the STAPLE analysis.With regard to our sector specific analysis, the sectors that exhibit thelowest agreement between model and manual contours, namely the midg-land and apical posterior (MPS and APS), are regions where the probe warpdeformation has its greatest effect. The fact that model-generated contourstend to ?over-contour? this sector may indicate that the unwarping proce-dure requires modification, ie. the posterior may need to be ?flattened?in order to resemble manual CT contours. The original intent of the un-warping algorithm was to facilitate 2-D ellipse-fitting to the axial contours[4], not necessarily to reproduce prostate anatomy, so it is perhaps unsur-prising that the algorithm would require adjustment. However, it is also47Chapter 4. Discussionimportant to keep in mind that our reference for comparison, ie. manualCT contouring, is itself subject to large uncertainties. Smith et al. [73]analysed manually drawn contours on 3-dimensional TRUS (3DTRUS), CTand MR images and found that CT contouring showed the greatest degreeof inter- and intra-observer variability. They also report that MR contoursat the apex were typically as large as, or larger than, the base, suggestingthat perhaps prostate anatomy does not taper in the inferior direction asdrastically as interpreted on other types of imaging. Additionally, Gao et al.[27] report that CT contours by radiation oncologists tend to systematicallyomit some of the prostate posterior, while including anterior normal tissue,when compared with photographic images of a human cadaver. In the fu-ture, a more in-depth analysis of model-based contouring might include MRimaging for validation purposes. It would also be interesting to generatea set of physician-reviewed model-based contours, similar to the analysisof Mahdavi et al. [41], and evaluate their level of acceptance by a groupof expert observers, compared against other anonymized, non-model-basedcontours.As a general observation, our data confirms previous findings in researchcarried out at this institution that the anterior superior quadrant (ASQ),corresponding to our BAS and MAS, is being routinely underdosed com-pared to other regions [70]. It is clear that V100 and D90 to the BAS, as seenin Figure A.1, is well below suboptimal cutoff values of 85% and 130 Gy,respectively, for the majority of patient cases. This is observed for the MASin Figure A.2 as well, but to a lesser degree. Poor dose coverage in theASQ may be attributed to a number of factors, including intentional avoid-ance of excessive dose in the urethra and bladder neck, needle drag or splayduring implantation, or contouring uncertainties [70]. However, dosimetricparameters for the whole prostate [50] or the ASQ [79] are not predictive ofbiochemical relapse in a large cohort study from the same institution. Thishas been observed in other investigations as well [eg. 10, 11], in contrastto the majority of reports demonstrating that dose response [eg. 63, 82], aswell as contouring and seed segmentation uncertainties [37], are tied to bi-ological outcomes. Reasons for this discrepancy, and the role of traditionaldose metrics in assessing treatment quality, are matters of ongoing debate[49, 81].Another notable observation is that TRUS-based model contours gener-ate larger volumes overall relative to manual CT contours. This is perhapsunexpected on first glance, since CT volumes are generally reported to belarger than TRUS (eg. postimplant CT/TRUS volume ratio 1.13 [77], 1.30[73]). However, as previously discussed, the model unwarping procedure484.1. Modeling as an Alternative to MR-CT Fusionadds volume to the raw TRUS contours, which likely accounts for some ofthe effect.Finally, Friedman testing further demonstrates the equivalence betweenmodel-based and manually-generated CT contours. From a statistical stand-point, both modeling methods generated contours that are no more differ-ent from the group as a whole than any individual observer. If prostatebrachytherapy treatment quality continues to be assessed mainly on dosi-metric parameters such as V100 and D90 that are derived from CT contours,an ultrasound-based model could provide some assistance in the contouringprocess, if not replace it outright. In a clinical setting, model-based contourscould be used as a starting point to guide physicians, who could then furthermodify and approve the final segmentation.4.1 Modeling as an Alternative to MR-CT FusionMR-CT fusion has long been considered a solution to the difficulties as-sociated with delineation of soft tissue on CT alone [64, 66]. Typically,fusion-based dosimetry is performed on Day-30 MR and CT images, whenthe effects of edematic expansion of the prostate have lessened [57]. Be-tween different imaging modalities, MR typically yields the least variability[19, 24, 73], although the medium itself is not without uncertainties. Forexample, De Brabandere et al. [22] report a ?surprisingly large? degree ofinterobserver contouring variability on T2-weighted MR images, and sug-gest that the fusion step itself may also be a ?weak link? in the procedure.Possible disadvantages aside, MR imaging is not routinely available for allprostate brachytherapy patients, and to obtain it requires extra time and ef-fort on behalf of the patient, not to mention the additional cost. Ultrasound,on the other hand, is regularly utilized for intraoperative guidance duringprostate brachytherapy implants, and yields comparable contouring results.Smith et al. [73] report that in volume, length and variability, prostatecontouring on 3DTRUS is most consistent with MR. In a study consideringMR images for preimplant treatment planning, Liu et al. [38] find strongsimilarities between US and MR (volume ratio 0.99? 0.08), with small dif-ferences in volume and dimension being significant intraobserver, but notinterobserver. Greater TRUS contour variability observed postimplant (eg.13% median standard deviation of the volume vs. 7% preimplant [72]) maybe attributed to the presence of seeds or intraprostatic hemorrhage obscur-ing prostate boundaries [92]. Thus, the early intraoperative TRUS imagesobtained in this work (after ?19% of seeds have been implanted, on average)494.2. Seed Localization in TRUSmight be spared some of the seed-induced confusion.The correspondence between MR and TRUS therefore suggests that con-tours derived from TRUS-based modeling could provide a suitable substitutefor the often-cited ?gold standard? of MR-CT fusion, when such resourcesare not readily available.This is not the first study to explore the option of incorporating TRUSinto postoperative dosimetry. Xue et al. [92] demonstrated the feasibilityof dosimetry performed on TRUS directly by segmenting both the prostateand seeds directly on TRUS, but found that interobserver contour variabil-ity on postimplant TRUS had a prohibitive effect on the calculated V100 andD90, showing variations that are less than, but approach, those observed onCT. A follow-up statistical analysis, however, found TRUS and CT-baseddosimetry to be indistinguishable [14]. Others [8, 44, 55] have fused preim-plant TRUS to postimplant CT based on urethral or rectal anatomy andfound comparable dosimetric parameters to MR-CT fusion; however, probedeformation and volume effects such as edema are described as some ofthe limitations of this technique. Many groups have reported on combiningTRUS with fluoroscopy and shown promising results, for the purposes of in-traoperative planning [23, 48, 86] or dosimetry [78, 83]. In the present study,our approach using mathematical modeling based on intraoperative TRUS isnovel and comes with many benefits. Prostate edge detection is less suscep-tible to user bias and image degradation caused by ?100 seeds postimplant.Our model-based procedure eliminates the need for manual image fusion,and uses intraoperative, rather than preoperative, TRUS. Registration isbased on matched seed locations, rather than anatomical features that maypotentially be poorly defined, such as the urethra without catheterization.Furthermore, utilization of a model allows deformations such as probe warp-ing and edema to be taken into account, and possibly corrected for. Finally,we reiterate that truly only resources that are currently in place are needed,eliminating any inconvenience to the patient (ie. fiducial markers, specialimaging equipment, urethrography, or postoperative TRUS volume studyare not required).4.2 Seed Localization in TRUSOne of the main limitations to using ultrasound for model-based CT contour-ing is in the ability to accurately calculate a coordinate registration throughmatched seeds. Seed localization and identification on CT is very accuratethanks to in-house software [15], the results of which are then manually504.2. Seed Localization in TRUSverified, but on TRUS this is not a straightforward task [30]. Regardless,ultrasound seed localization is essential to the methods of this algorithm,and we have attempted to quantify the uncertainty, estimating ?0.24 cmerror in seed locations.A variety of factors contribute to the uncertainty. Most notably, brachy-therapy seeds on ultrasound are difficult to distinguish from other objectsexhibiting similar echogenicity, such as the connecting brachytherapy strandmaterial or soft tissue calcifications, leading to confusion and detection offalse positives and false negatives. The high acoustic reflectivity (specular-ity) of seed casings also creates shadows more distal from the transducerprobe, and rotation of cylindrical seeds out of the implant axis can renderthem ?invisible? on a transverse scan. Despite these challenges, significanteffort has been directed towards development of algorithms to automaticallydetect seeds in TRUS (including here at UBC), with preliminary resultsshowing success in controlled test environments [eg. 43, 88, 89, 93]. Otherauthors have reported on the accuracy of performing manual seed segmen-tation, with varying levels of success ranging from full reconstruction [91]to only 20%?25% of seeds being identified [60]. For our purposes, withoutaccess to the tools utilized by some groups, we are limited to data and tech-niques that are clinically available, namely manual identification on axialB-mode images. Recall however that with only a partial implant in place,many of the problems encountered when attempting a full postimplant re-construction, such as confusion or shadowing, are alleviated. In addition,the use of sagittal imaging allowed better visualization of patient anatomy,particularly at the base and apex. Future work may potentially include theuse of more advanced sagittal ultrasound techniques, such as those describedby Wen et al. [89].The use of flexible ThinStrands may have compounded some of the neg-ative effects in 6 cases: implanted sources appeared to undergo increasedrotation and displacement, and this, combined with their smaller diameters,may have lead to greater difficulty in detection for both ultrasound and flu-oroscopy. This is despite studies on phantoms examining multiple imagingplatforms that have reported that thin seeds are as easily detected (and pos-sibly even more so) as standard thickness seeds [67, 68]. However, perhapssome of the stated advantages were lost when applied to our clinical patientdata, in contrast to other findings [84], possibly due to inexperience on thepart of the implanting physician.Even when seeds are adequately visualized, sparsely sampled TRUS slices(0.5 cm separation), leading to a high probability of seeds falling betweenslices or beyond the base or apex slice, hinders accurate identification of514.2. Seed Localization in TRUSthe correct strand and seed number. Perhaps future studies should includecollection of a larger volume TRUS study, particularly extending inferior tothe prostate apex. Despite this, it was possible to locate ?93% of seeds onaverage, mainly because of our knowledge of the seeds present at the timeof imaging, and our reliance on fluoroscopic visualization.52Chapter 5ConclusionsPostimplant dosimetry is of utmost importance for the continuous assuranceof treatment quality to patients undergoing brachytherapy for the treat-ment of prostate cancer. However, dosimetry based on manually-drawnCT contours, the current standard of care, is known to suffer from largeuncertainties due to the considerable intra- and interobserver variability.Clearly, there is a need for methods that can improve the reliability of cal-culated dose metrics. To that aim, we have demonstrated the feasibilityof using ultrasound-based contours to model patient anatomy and producecontours for CT-based dosimetry. This was possible due to our ability toperform a full source plan reconstruction on postoperative CT images, al-lowing a matched seed-to-seed correspondence between TRUS and CT to beobtained, from which a coordinate registration could be calculated. Math-ematical modeling, based on a modified 3-dimensional ellipsoid shape, per-mits anatomy as seen on one imaging modality (TRUS) to be deformed tobetter represent its appearance on another (CT).In general, model-based prostate contours are within the variability rangeobserved in a set of five manual contours, all generated by experienced physi-cians well versed in prostate contouring on CT images. We evaluate the qual-ity of model-generated contouring in several ways. In our analysis based onthe STAPLE algorithm and definitions of sensitivity and positive predictivevalue, we find that model-generated contours select all the voxels that mostlikely represent true structure (ie. the sensitivity of the model, or recall)with the same frequency as manual contours, in nearly all cases (?19 outof 20). However, modeling also selected more voxels that are likely non-prostate structure (ie. it has a lower positive predictive value, or precision)than manual contours in over half of the cases.Volume and dosimetric analysis reveals that modeling tends to predict atotal prostate volume that is slightly greater than, but still statistically con-sistent with, manual contouring. This increased volume, however, translatesto less agreement of calculated dose parameters V100 and D90, and motivatesour region-specific analysis. By dividing the prostate volume into 9 sectors,we are able to determine that the posterior sectors of the apex and midgland53Chapter 5. Conclusionsare the most responsible for this dosimetric disagreement, with the modeltending to over-contour these regions relative to the mean manual value,leading to an underestimate of the dose metrics in these sectors. We hy-pothesize that these sectors, corresponding to the region most affected byTRUS probe compression, may imply that our algorithm to ?unwarp? theanatomical deformation requires further refinement. However, our standardfor comparison is itself subject to uncertainty, which must be factored intoany interpretation of the results.Statistical testing using the Friedman test for repeated-measures, non-parametric data indicates that, based on volumetric and dosimetric param-eters, the contours generated by our 7 total ?observers? are not equivalent.Subsequent post-hoc analysis reveals that model-based contours are no moredifferent from the group as a whole than any individual manual-contour-generator, at a significance level of ? = 0.05. From a purely statisticalstandpoint, this seems to imply that dose metrics derived directly frommodel-based contours are no more ?uncertain? than those from manuallydrawn contours; however, our knowledge of the spatial geometry suggestscaution should be exercised before drawing such a conclusion.There are several important benefits to creating CT contours from amodel that is based on TRUS images. Contours are generated in a consistentway using semi-automatic algorithms, so they are less subject to individualbias or confusion from image artifacts, such as those originating from theseeds themselves. Another significant strength of our model-based methodis that only resources that are already available in the clinic are required,which has the obvious benefit of being convenient and noninvasive for thepatient, as well as cost-effective from a public health standpoint. AlthoughMR is frequently cited as being the ?gold standard? for postimplant prostatedosimetry, strong similarities between anatomical information retrieved fromMR and ultrasound images suggests that a TRUS-based-model may providea suitable alternative, in cases where MR is not an option. Also, comparedto methods that rely on raw TRUS images fused to CT, our techniqueinvolving a deformable model and seed-based registration may produce abetter representation of true patient anatomy.Several limitations of this method do exist, the most significant being theongoing challenge of seed localization on TRUS. Seeds are notoriously diffi-cult to see on B-mode ultrasound images, even with careful manual inspec-tion. Therefore, any attempt to automate this procedure would prove dif-ficult, if not impossible, without the introduction of specialized equipment,a step that would compromise the ease-of-use and ?readily available? assetsof this technique. Although there are currently other manually-performed54Chapter 5. Conclusionssteps, such as the adjustment of axial images to match sagittal boundaries,these could likely be automated with a moderate amount of technical ma-nipulation.If ultrasound seed localization could be performed with relative ease andon a reasonable timescale, we propose that model-based contours could beapplied in a clinical setting. It is likely that manual modification, or atleast review, by a physician would be recommended, however, the techniqueof using semi-automatic contouring as a starting point for physicians is notuncommon, and in fact, is the basis for routine treatment planning practicesat some centers in the BCCA.In future studies, it would be advantageous to adapt the sagittal ultra-sound image collection procedure (eg. perhaps obtaining additional peris-agittal images, in addition to the midsagittal) to provide better visualizationof organ boundaries, and to devise a more accurate registration method be-tween sagittal and transverse ultrasound. To capture more extraprostaticseeds, collecting TRUS image slices inferior to the apex slice of a standardvolume study would be simple to execute and immensely useful. A refinedunwarping algorithm to modify the prostate model posterior also warrantsexamination, guided perhaps by information on prostate anatomy providedby surgical research. Finally, access to MR images of the patient cohort forvalidation purposes would simplify, and perhaps improve, the evaluation ofthe model algorithm.55Bibliography[1] Source: BC Cancer Agency. Brachytherapy Guidelines. Accessed July 2013.[2] Source: BC Cancer Agency. Prostate. Accessed July 2013.[3] R. J. Amdur, D. Gladstone, K. A. Leopold, and R. D. Harris. Prostateseed implant quality assessment using MR and CT image fusion. Int.J. Radiat. Oncol. Biol. Phys., 43(1):67?72, 1999.[4] S. Badiei. Prostate segmentation in ultrasound images using imagewarping and ellipsoid fitting. Master?s thesis, University of BritishColumbia, 2007.[5] V. Bewick, L. Cheek, and J. Ball. Statistics review 10: further non-parametric methods. Crit Care, 8(3):196?199, 2004.[6] W. S. Bice, B. R. Prestidge, P. D. Grimm, J. L. Friedland, V. Feygel-man, M. Roach, J. J. Prete, D. F. Dubois, and J. C. Blasko. Cen-tralized multiinstitutional postimplant analysis for interstitial prostatebrachytherapy. Int. J. Radiat. Oncol. Biol. Phys., 41(4):921?927, 1998.[7] J. C. Blasko, T. Mate, J. E. Sylvester, P. D. Grimm, and W. Cavanagh.Brachytherapy for carcinoma of the prostate: techniques, patient selec-tion, and clinical outcomes. Semin Radiat Oncol, 12(1):81?94, 2002.[8] D. Bowes, J. M. Crook, C. Araujo, and D. Batchelar. Ultrasound-CT fusion compared with MR-CT fusion for postimplant dosimetry inpermanent prostate brachytherapy. Brachytherapy, 12(1):38?43, 2013.[9] D. Bowes, J. M. Crook, R. Rajapakshe, C. Araujo, and B. Parker. Defin-ing a magnetic resonance scan sequence for permanent seed prostatebrachytherapy postimplant assessment. Brachytherapy, 12(1):25?29,2013.56Bibliography[10] A. P. Brown, T. J. Pugh, D. A. Swanson, R. J. Kudchadker, T. L.Bruno, E. N. Christensen, M. van Vulpen, and S. J. Frank. Improvingprostate brachytherapy quality assurance with MRI-CT fusion-basedsector analysis in a phase II prospective trial of men with intermediate-risk prostate cancer. Brachytherapy, 2013. In press.[11] W. M. Butler, R. R. Stewart, and G. S. Merrick. A detailed radio-biological and dosimetric analysis of biochemical outcomes in a case-control study of permanent prostate brachytherapy patients. Med Phys,36(3):776?787, 2009.[12] Source: Prostate Cancer Canada. About Prostate Can-cer. Accessed July 2013.[13] J. H. Challis. A procedure for determining rigid body transformationparameters. J Biomech, 28(6):733?737, 1995.[14] M. S. Chew, J. Xue, C. Houser, V. Misic, J. Cao, T. Cornwell, J. Han-dler, Y. Yu, and E. Gressen. Impact of transrectal ultrasound- andcomputed tomography-based seed localization on postimplant dosime-try in prostate brachytherapy. Brachytherapy, 8(2):255?264, 2009.[15] N. Chng, I. Spadinger, W. J. Morris, N. Usmani, and S. Salcudean.Prostate brachytherapy postimplant dosimetry: automatic plan recon-struction of stranded implants. Med Phys, 38(1):327?342, 2011.[16] N. T. Chng. Characterization of Implanted Seed Orientation and Dis-placement Dynamics with Application to the Design of Non-UniformSource Strength Treatment Plans for Prostate Brachytherapy. PhD the-sis, University of British Columbia, 2012.[17] P. P. Connell and S. Hellman. Advances in radiotherapy and impli-cations for the next century: a historical perspective. Cancer Res.,69(2):383?392, 2009.[18] W. J. Conover. Practical Nonparametric Statistics. John Wiley & SonsInc., New York, 1999.[19] J. Crook, M. Milosevic, P. Catton, I. Yeung, T. Haycocks, T. Tran,C. Catton, M. McLean, T. Panzarella, and M. A. Haider. Interobservervariation in postimplant computed tomography contouring affects qual-ity assessment of prostate brachytherapy. Brachytherapy, 1(2):66?73,2002.57Bibliography[20] A. V. D?Amico, R. Whittington, S. B. Malkowicz, D. Schultz, K. Blank,G. A. Broderick, J. E. Tomaszewski, A. A. Renshaw, I. Kaplan, C. J.Beard, and A. Wein. Biochemical outcome after radical prostatectomy,external beam radiation therapy, or interstitial radiation therapy forclinically localized prostate cancer. JAMA, 280(11):969?974, 1998.[21] B. J. Davis, E. M. Horwitz, W. R. Lee, J. M. Crook, R. G. Stock, G. S.Merrick, W. M. Butler, P. D. Grimm, N. N. Stone, L. Potters, A. L.Zietman, and M. J. Zelefsky. American Brachytherapy Society con-sensus guidelines for transrectal ultrasound-guided permanent prostatebrachytherapy. Brachytherapy, 11(1):6?19, 2012.[22] M. De Brabandere, P. Hoskin, K. Haustermans, F. Van den Heuvel, andF. A. Siebert. Prostate post-implant dosimetry: interobserver vari-ability in seed localisation, contouring and fusion. Radiother Oncol,104(2):192?198, 2012.[23] E. Dehghan, J. Lee, P. Fallavollita, N. Kuo, A. Deguet, Y. Le,E. Clif Burdette, D. Y. Song, J. L. Prince, and G. Fichtinger.Ultrasound-fluoroscopy registration for prostate brachytherapy dosime-try. Med Image Anal, 16(7):1347?1358, 2012.[24] D. F. Dubois, B. R. Prestidge, L. A. Hotchkiss, J. J. Prete, and W. S.Bice. Intraobserver and interobserver variability of MR imaging- andCT-derived prostate volumes after transperineal interstitial permanentprostate brachytherapy. Radiology, 207(3):785?789, 1998.[25] Source: Mayo Foundation for Medical Education and Research. Perma-nent prostate brachytherapy. Accessed July 2013.[26] D. B. Fuller and H. Jin. Computed tomography-ultrasound fusionbrachytherapy: description and evolution of the technique. Brachyther-apy, 6(4):272?279, 2007.[27] Z. Gao, D. Wilkins, L. Eapen, C. Morash, Y. Wassef, and L. Gerig. Astudy of prostate delineation referenced against a gold standard createdfrom the visible human data. Radiother Oncol, 85(2):239?246, 2007.[28] L. G. Gomella, X. S. Liu, E. J. Trabulsi, W. K. Kelly, R. Myers,T. Showalter, A. Dicker, and R. Wender. Screening for prostate can-cer: the current evidence and guidelines controversy. Can J Urol,18(5):5875?5883, 2011.58Bibliography[29] B. H. Han, K. Wallner, G. Merrick, K. Badiozamani, and W. But-ler. The effect of interobserver differences in post-implant prostate CTimage interpretation on dosimetric parameters. Med Phys, 30(6):1096?1102, 2003.[30] B. H. Han, K. Wallner, G. Merrick, W. Butler, S. Sutlief, andJ. Sylvester. Prostate brachytherapy seed identification on post-implantTRUS images. Med Phys, 30(5):898?900, 2003.[31] M. Hollander and D. A. Wolfe. Nonparametric Statistical Methods. JohnWiley & Sons Inc., New York, 1973.[32] H. H. Holm, N. Juul, J. F. Pedersen, H. Hansen, and I. Str?yer.Transperineal 125iodine seed implantation in prostatic cancer guidedby transrectal ultrasonography. J. Urol., 130(2):283?286, 1983.[33] Source: National Cancer Institute. Fact Sheet on Radiation Ther-apy for Cancer. Accessed July 2013.[34] Source: National Cancer Institute. Prostate and Nearby Or-gans. Accessed July 2013.[35] M. Keyes, W. J. Morris, I. Spadinger, C. Araujo, A. Cheung, N. Chng,J. Crook, R. Halperin, V. Lapointe, S. Miller, H. Pai, and T. Pickles.Radiation oncology and medical physicists quality assurance in BritishColumbia Cancer Agency Provincial Prostate Brachytherapy Program.Brachytherapy, 12(4):343?355, 2013.[36] L. Klotz. Active surveillance with selective delayed intervention forfavorable risk prostate cancer. Urol. Oncol., 24(1):46?50, 2006.[37] P. E. Lindsay, J. Van Dyk, and J. J. Battista. A systematic study ofimaging uncertainties and their impact on 125I prostate brachytherapydose evaluation. Med Phys, 30(7):1897?1908, 2003.[38] D. Liu, N. Usmani, S. Ghosh, W. Kamal, J. Pedersen, N. Pervez, D. Yee,B. Danielson, A. Murtha, J. Amanie, and R. S. Sloboda. Compari-son of prostate volume, shape, and contouring variability determinedfrom preimplant magnetic resonance and transrectal ultrasound images.Brachytherapy, 11(4):284?291, 2012.59Bibliography[39] J. Lobo. Towards intra-operative dosimetry for prostate brachytherapy:improved seed detection and registration to ultrasound using needledetection. Master?s thesis, University of British Columbia, 2012.[40] S. S. Mahdavi, N. Chng, I. Spadinger, W. J. Morris, and S. E. Sal-cudean. Semi-automatic segmentation for prostate interventions. MedImage Anal, 15(2):226?237, 2011.[41] S. S. Mahdavi, I. Spadinger, N. Chng, S. E. Salcudean, and W. J.Morris. Semiautomatic segmentation for prostate brachytherapy: dosi-metric evaluation. Brachytherapy, 12(1):65?76, 2013.[42] K. L. Maletz, R. D. Ennis, J. Ostenson, A. Pevsner, A. Kagen, andI. Wernick. Comparison of CT and MR-CT fusion for prostate post-implant dosimetry. Int. J. Radiat. Oncol. Biol. Phys., 82(5):1912?1917,2012.[43] J. Mamou, S. Ramachandran, and E. J. Feleppa. Angle-dependentultrasonic detection and imaging of brachytherapy seeds using singularspectrum analysis. J. Acoust. Soc. Am., 123(4):2148?2159, 2008.[44] L. G. Marcu, J. M. Lawson, T. Rutten, and R. Gowda. Quality in-dicators and technique for analyzing permanent I-125 prostate seedimplants: seven years postimplant dosimetry evaluation. Med Phys,39(7):4123?4131, 2012.[45] Source: Urology Match. I. Basic Principles: Prostate Anatomy. Accessed July2013.[46] P. W. McLaughlin, C. Evans, M. Feng, and V. Narayana. Radiographicand anatomic basis for prostate contouring errors and methods to im-prove prostate contouring accuracy. Int. J. Radiat. Oncol. Biol. Phys.,76(2):369?378, 2010.[47] P. W. McLaughlin, S. Troyer, S. Berri, V. Narayana, A. Meirowitz,P. L. Roberson, and J. Montie. Functional anatomy of the prostate:implications for treatment planning. Int. J. Radiat. Oncol. Biol. Phys.,63(2):479?491, 2005.[48] M. Moradi, S. S. Mahdavi, E. Dehghan, J. R. Lobo, S. Deshmukh,W. J. Morris, G. Fichtinger, and S. T. Salcudean. Seed localizationin ultrasound and registration to C-arm fluoroscopy using matched60Bibliographyneedle tracks for prostate brachytherapy. IEEE Trans Biomed Eng,59(9):2558?2567, 2012.[49] W. J. Morris, R. Halperin, and I. Spadinger. Point: the relationship be-tween postimplant dose metrics and biochemical no evidence of diseasefollowing low dose rate prostate brachytherapy: is there an elephant inthe room? Brachytherapy, 9(4):289?292, 2010.[50] W. J. Morris, M. Keyes, D. Palma, M. McKenzie, I. Spadinger, A. Agra-novich, T. Pickles, M. Liu, W. Kwan, J. Wu, V. Lapointe, E. Berthelet,H. Pai, R. Harrison, W. Kwa, J. Bucci, V. Racz, and R. Woods. Evalu-ation of dosimetric parameters and disease response after 125 iodinetransperineal brachytherapy for low- and intermediate-risk prostatecancer. Int. J. Radiat. Oncol. Biol. Phys., 73(5):1432?1438, 2009.[51] W. J. Morris, M. Keyes, I. Spadinger, W. Kwan, M. Liu, M. McKen-zie, H. Pai, T. Pickles, and S. Tyldesley. Population-based 10-yearoncologic outcomes after low-dose-rate brachytherapy for low-risk andintermediate-risk prostate cancer. Cancer, 119(8):1537?1546, 2013.[52] S. Nag, D. Beyer, J. Friedland, P. Grimm, and R. Nath. AmericanBrachytherapy Society (ABS) recommendations for transperineal per-manent brachytherapy of prostate cancer. Int. J. Radiat. Oncol. Biol.Phys., 44(4):789?799, 1999.[53] S. Nag, W. Bice, K. DeWyngaert, B. Prestidge, R. Stock, and Y. Yu.The American Brachytherapy Society recommendations for permanentprostate brachytherapy postimplant dosimetric analysis. Int. J. Radiat.Oncol. Biol. Phys., 46(1):221?230, 2000.[54] V. Narayana, P. L. Roberson, A. T. Pu, H. Sandler, R. H. Winfield, andP. W. McLaughlin. Impact of differences in ultrasound and computedtomography volumes on treatment planning of permanent prostate im-plants. Int. J. Radiat. Oncol. Biol. Phys., 37(5):1181?1185, 1997.[55] V. Narayana, P. L. Roberson, R. J. Winfield, and P. W. McLaughlin.Impact of ultrasound and computed tomography prostate volume reg-istration on evaluation of permanent prostate implants. Int. J. Radiat.Oncol. Biol. Phys., 39(2):341?346, 1997.[56] R. Nath, L. L. Anderson, G. Luxton, K. A. Weaver, J. F. Williamson,and A. S. Meigooni. Dosimetry of interstitial brachytherapy sources:recommendations of the AAPM Radiation Therapy Committee Task61BibliographyGroup No. 43. American Association of Physicists in Medicine. MedPhys, 22(2):209?234, 1995.[57] R. Nath, W. S. Bice, W. M. Butler, Z. Chen, A. S. Meigooni,V. Narayana, M. J. Rivard, and Y. Yu. AAPM recommendationson dose prescription and reporting methods for permanent interstitialbrachytherapy for prostate cancer, Dec 2009. Full Report of AAPMTask Group 137.[58] R. Nath, W. S. Bice, W. M. Butler, Z. Chen, A. S. Meigooni,V. Narayana, M. J. Rivard, and Y. Yu. AAPM recommendationson dose prescription and reporting methods for permanent interstitialbrachytherapy for prostate cancer: report of Task Group 137. MedPhys, 36(11):5310?5322, 2009. Executive Summary.[59] Oncura Inc., 3350 North Ridge Avenue, Arlington Heights, IL 60004USA. Instructions for the use of RAPID Strand for InterstitialBrachytherapy Treatments, 2003. Revised August 2009.[60] P. F. Orio, I. B. Tutar, S. Narayanan, S. Arthurs, P. S. Cho, Y. Kim,G. Merrick, and K. E. Wallner. Intraoperative ultrasound-fluoroscopyfusion can enhance prostate brachytherapy quality. Int. J. Radiat. On-col. Biol. Phys., 69(1):302?307, 2007.[61] T. Pickles. What?s a man to do? Treatment options for localizedprostate cancer. Can Fam Physician, 50:65?72, 2004.[62] A. Polo, F. Cattani, A. Vavassori, D. Origgi, G. Villa, H. Marsiglia,M. Bellomi, G. Tosi, O. De Cobelli, and R. Orecchia. MR and CT imagefusion for postimplant analysis in permanent prostate seed implants.Int. J. Radiat. Oncol. Biol. Phys., 60(5):1572?1579, 2004.[63] L. Potters, D. Huang, E. Calugaru, P. Fearn, L. Lee, and M. W. Kat-tan. Importance of implant dosimetry for patients undergoing prostatebrachytherapy. Urology, 62(6):1073?1077, 2003.[64] C. Rasch, I. Barillot, P. Remeijer, A. Touw, M. van Herk, and J. V.Lebesque. Definition of the prostate in CT and MRI: a multi-observerstudy. Int. J. Radiat. Oncol. Biol. Phys., 43(1):57?66, 1999.[65] M. J. Rivard, B. M. Coursey, L. A. DeWerd, W. F. Hanson, M. S. Huq,G. S. Ibbott, M. G. Mitch, R. Nath, and J. F. Williamson. Updateof AAPM Task Group No. 43 Report: A revised AAPM protocol forbrachytherapy dose calculations. Med Phys, 31(3):633?674, 2004.62Bibliography[66] M. Roach, P. Faillace-Akazawa, C. Malfatti, J. Holland, and H. Hricak.Prostate volumes defined by magnetic resonance imaging and comput-erized tomographic scans for three-dimensional conformal radiotherapy.Int. J. Radiat. Oncol. Biol. Phys., 35(5):1011?1018, 1996.[67] G. Roberts, B. Al-Qaisieh, and P. Bownes. Evaluation of the visi-bility of a new thinner 125I radioactive source for permanent prostatebrachytherapy. Brachytherapy, 11(6):460?467, 2012.[68] A. K. Robertson, P. S. Basran, S. D. Thomas, and D. Wells. CT,MR, and ultrasound image artifacts from prostate brachytherapy seedimplants: the impact of seed size. Med Phys, 39(4):2061?2068, 2012.[69] S. A. Rosenthal, N. H. Bittner, D. C. Beyer, D. J. Demanes, B. J.Goldsmith, E. M. Horwitz, G. S. Ibbott, W. R. Lee, S. Nag, W. W.Suh, and L. Potters. American Society for Radiation Oncology (AS-TRO) and American College of Radiology (ACR) practice guideline forthe transperineal permanent brachytherapy of prostate cancer. Int. J.Radiat. Oncol. Biol. Phys., 79(2):335?341, 2011.[70] S. Sidhu, W. J. Morris, I. Spadinger, M. Keyes, M. Hilts, R. Harrison,K. Otto, M. McKenzie, and A. Agranovich. Prostate brachytherapypostimplant dosimetry: a comparison of prostate quadrants. Int. J.Radiat. Oncol. Biol. Phys., 52(2):544?552, 2002.[71] R. S. Sloboda, N. Usmani, J. Pedersen, A. Murtha, N. Pervez, andD. Yee. Time course of prostatic edema post permanent seed implantdetermined by magnetic resonance imaging. Brachytherapy, 9(4):354?361, 2010.[72] S. Smith, K. Wallner, G. Merrick, W. Butler, S. Sutlief, and P. Grimm.Interpretation of pre- versus postimplant trus images. Med Phys,30(5):920?924, 2003.[73] W. L. Smith, C. Lewis, G. Bauman, G. Rodrigues, D. D?Souza, R. Ash,D. Ho, V. Venkatesan, D. Downey, and A. Fenster. Prostate volumecontouring: a 3D analysis of segmentation using 3DTRUS, CT, andMR. Int. J. Radiat. Oncol. Biol. Phys., 67(4):1238?1247, 2007.[74] Source: American Brachytherapy Society. A Brief Essay on the In-troduction of Brachytherapy. Accessed July 2013.63Bibliography[75] Source: Canadian Cancer Society. How radiation therapyworks. Accessed July 2013.[76] Source: Canadian Cancer Society. Prostate-specific antigen(PSA) test. Accessed July 2013.[77] M. C. Solhjem, B. J. Davis, T. M. Pisansky, T. M. Wilson, L. A. Myn-derse, M. G. Herman, B. F. King, and S. M Geyer. Prostate volumemeasurement by transrectal ultrasound and computed tomography be-fore and after permanent prostate brachytherapy. Int. J. Radiat. Oncol.Biol. Phys., 60(3):767?776, 2004.[78] D. Y. Song, A. K. Jain, Z. Zhang, A. Deguet, Y. Le, E. Armour, E. C.Burdette, and G. Fichtinger. Dynamic intraoperative dosimetry forprostate brachytherapy using a nonisocentric C-arm. Brachytherapy,10(2):98?106, 2011.[79] I. Spadinger, W. J. Morris, M. Keyes, M. Liu, R. Shaffer, A. Thompson,N. Chng, and R. Woods. Quadrant dosimetry as a predictor of biochem-ical relapse in 125I prostate brachytherapy. Brachytherapy, 10(2):87?97,2011.[80] M. Steggerda, C. Schneider, M. van Herk, L. Zijp, L. Moonen, andH. van der Poel. The applicability of simultaneous TRUS-CT imagingfor the evaluation of prostate seed implants. Med Phys, 32(7):2262?2270, 2005.[81] R. G. Stock. Counterpoint: there is a dose-response relationshipin the low-dose rate brachytherapy management of prostate cancer.Brachytherapy, 9(4):293?296, 2010.[82] R. G. Stock, N. N. Stone, A. Tabert, C. Iannuzzi, and J. K. DeWyn-gaert. A dose-response study for I-125 prostate implants. Int. J. Radiat.Oncol. Biol. Phys., 41(1):101?108, 1998.[83] Y. Su, B. J. Davis, K. M. Furutani, M. G. Herman, and R. A.Robb. Seed localization and TRUS-fluoroscopy fusion for intraoperativeprostate brachytherapy dosimetry. Comput. Aided Surg., 12(1):25?34,2007.64Bibliography[84] J. Sylvester, P. Grimm, D. Naidoo, J. Bilik, A. Miller, and J. Wong.First report on the use of a thinner (125)I radioactive seed within 20-gauge needles for permanent radioactive seed prostate brachytherapy:Evaluation of postimplant dosimetry and acute toxicity. Brachytherapy,12(4):375?381, 2013.[85] O. Tanaka, S. Hayashi, K. Sakurai, M. Matsuo, M. Nakano, S. Maeda,H. Hoshi, and T. Deguchi. Importance of the CT/MRI fusion methodas a learning tool for CT-based postimplant dosimetry in prostatebrachytherapy. Radiother Oncol, 81(3):303?308, 2006.[86] I. B. Tutar, L. Gong, S. Narayanan, S. D. Pathak, P. S. Cho, K. Wallner,and Y. Kim. Seed-based transrectal ultrasound-fluoroscopy registrationmethod for intraoperative dosimetry analysis of prostate brachytherapy.Med Phys, 35(3):840?848, 2008.[87] S. K. Warfield, K. H. Zou, and W. M. Wells. Simultaneous truth andperformance level estimation (STAPLE): an algorithm for the valida-tion of image segmentation. IEEE Trans Med Imaging, 23(7):903?921,2004.[88] Z. Wei, L. Gardi, D. B. Downey, and A. Fenster. Automated localizationof implanted seeds in 3D TRUS images used for prostate brachytherapy.Med Phys, 33(7):2404?2417, 2006.[89] X. Wen, S. T. Salcudean, and P. D. Lawrence. Detection of brachyther-apy seeds using 3-D transrectal ultrasound. IEEE Trans Biomed Eng,57(10):2467?2477, 2010.[90] H. Westendorp, R. Kattevilder, A. van ?t Riet, A. W. Minken, T. T.Nuver, J. J. Immerzeel, and C. J. Hoekstra. Objective automated as-sessment of time trends in prostate edema after (125)I implantation.Brachytherapy, 11(5):327?333, 2012.[91] J. Xue, F. Waterman, J. Handler, and E. Gressen. Localization of linked125I seeds in postimplant TRUS images for prostate brachytherapydosimetry. Int. J. Radiat. Oncol. Biol. Phys., 62(3):912?919, 2005.[92] J. Xue, F. Waterman, J. Handler, and E. Gressen. The effect of inter-observer variability on transrectal ultrasonography-based postimplantdosimetry. Brachytherapy, 5(3):174?182, 2006.65[93] Y. Yu, S. T. Acton, and K. Thornton. Detection of radioactive seedsin ultrasound images of the prostate. In Proc. IEEE Int. Conf. ImageProcess., volume 2, pages 319?322. IEEE, Oct 2001.[94] Y. Yu, L. L. Anderson, Z. Li, D. E. Mellenberg, R. Nath, M. C. Schell,F. M. Waterman, A. Wu, and J. C. Blasko. Permanent prostate seedimplant brachytherapy: report of the American Association of Physi-cists in Medicine Task Group No. 64. Med Phys, 26(10):2054?2076,1999.66Appendix AData Tables and FiguresPatient A B C D E X Y1 29.33 35.53 19.59 27.18 29.87 26.17 25.572 48.47 43.76 32.02 45.50 49.22 40.59 39.603 39.35 36.86 28.54 41.14 40.91 31.40 29.544 46.88 50.93 21.74 44.78 47.03 47.42 46.135 48.44 51.86 15.97 50.23 38.61 53.97 51.806 43.08 37.80 16.79 39.14 33.23 23.69 22.687 57.69 59.24 28.14 64.00 49.65 53.03 49.808 30.29 31.75 19.95 35.51 32.79 29.37 28.199 38.92 36.28 20.40 45.28 39.70 45.11 43.2810 39.23 33.55 17.15 43.27 37.38 35.15 33.2511 55.67 57.31 29.47 68.86 63.76 66.45 63.5512 33.20 21.31 9.70 27.52 21.15 23.33 21.5013 43.52 43.87 27.76 49.40 41.19 46.88 44.3714 45.03 48.19 28.00 59.98 53.87 55.12 51.6515 33.66 44.10 26.84 52.00 50.01 64.61 60.8116 38.31 45.86 24.21 46.94 38.19 48.05 43.9717 35.43 45.39 20.41 47.07 44.68 48.46 45.3518 48.38 47.63 30.54 53.51 52.61 58.05 53.9519 45.78 44.59 29.57 53.49 38.69 46.99 44.2020 43.86 32.43 22.40 47.58 37.77 33.11 31.32Table A.1: Prostate volume (cc) measured for all 20 patients, based on 5manual (A?E) and 2 model (X and Y) sets of contours.67Appendix A. Data Tables and FiguresPatient A B C D E X Y1 96.19 88.50 83.07 93.66 90.54 94.49 94.112 93.66 91.63 83.77 94.36 90.87 86.35 86.173 93.38 93.08 84.09 92.19 93.47 92.76 93.104 93.31 87.08 71.52 92.71 89.55 88.11 89.185 94.95 89.83 98.39 93.72 94.86 88.88 90.966 93.18 90.70 96.59 92.60 97.20 98.77 98.957 90.71 84.17 87.05 88.23 89.59 92.30 91.778 91.88 85.68 74.98 89.17 93.52 92.75 92.399 93.55 91.92 91.38 88.35 91.84 91.85 93.1110 98.54 95.38 97.54 96.04 97.65 95.26 95.5211 99.16 99.60 99.48 98.62 99.12 99.11 99.1512 97.54 99.39 99.77 98.11 99.95 96.05 97.4513 92.65 91.54 87.87 90.51 96.07 94.93 95.3814 99.05 98.18 99.77 98.74 97.99 96.91 97.9715 99.97 98.11 100.00 98.65 98.75 88.46 91.0816 98.50 95.65 88.34 93.37 94.67 92.15 93.5617 98.47 97.12 96.46 96.61 96.87 91.33 91.7218 90.74 90.11 90.49 93.52 92.90 94.77 96.5119 71.70 74.74 46.21 66.00 77.19 83.25 85.0820 91.72 94.34 86.76 85.83 92.71 95.92 97.61Table A.2: Prostate V100 (%) calculated for all 20 patients, based on 5manual (A?E) and 2 model (X and Y) sets of contours.68Appendix A. Data Tables and FiguresPatient A B C D E X Y1 159.94 138.76 123.91 154.60 146.23 157.07 156.962 156.12 149.29 122.88 156.09 146.59 130.20 129.503 152.32 152.48 129.65 149.11 153.20 150.24 151.404 154.46 137.09 103.46 152.27 142.55 137.65 141.535 161.01 143.66 167.18 156.97 162.40 140.52 147.336 151.55 146.66 157.86 150.67 161.47 170.80 172.067 145.56 130.45 138.71 140.63 143.26 149.06 148.418 151.54 128.50 98.22 141.07 158.40 156.01 154.549 155.68 151.41 148.35 137.87 151.50 150.53 154.9210 173.16 160.33 164.78 162.24 171.57 163.22 162.9811 176.29 173.90 171.88 169.91 174.62 174.26 175.1812 132.39 140.10 132.59 136.82 141.03 129.90 132.6113 150.65 148.10 139.57 145.05 160.39 156.44 158.0514 181.77 178.11 182.73 176.21 177.93 173.72 178.6015 180.30 174.37 178.72 177.52 175.95 137.09 149.5516 176.03 163.62 138.48 157.02 165.71 151.07 156.0717 183.10 174.85 164.10 173.03 172.85 149.23 150.9518 146.82 144.44 145.83 159.52 155.31 161.25 167.9319 110.86 117.98 74.92 100.13 123.25 133.69 136.2520 147.24 154.18 133.79 134.46 148.98 158.44 162.30Table A.3: Prostate D90 (Gy) calculated for all 20 patients, based on 5manual (A?E) and 2 model (X and Y) sets of contours.69Appendix A. Data Tables and FiguresPatient A B C D E X Y1 5 7 1 4 6 3 22 6 4 1 5 7 3 23 5 4 1 7 6 3 24 4 7 1 2 5 6 35 3 6 1 4 2 7 56 7 5 1 6 4 3 27 5 6 1 7 2 4 38 4 5 1 7 6 3 29 3 2 1 7 4 6 510 6 3 1 7 5 4 211 2 3 1 7 5 6 412 7 3 1 6 2 5 413 3 4 1 7 2 6 514 2 3 1 7 5 6 415 2 3 1 5 4 7 616 3 5 1 6 2 7 417 2 5 1 6 3 7 418 3 2 1 5 4 7 619 5 4 1 7 2 6 320 6 3 1 7 5 4 2Rj 83 84 20 119 81 103 70Table A.4: Ranks R(Xij) and rank sums Rj of prostate volume given inTable A.1, for use in the Friedman statistical test and post-hoc analysis.70Appendix A. Data Tables and FiguresPatient A B C D E X Y1 7 2 1 4 3 6 52 6 5 1 7 4 3 23 6 4 1 2 7 3 54 7 2 1 6 5 3 45 6 2 7 4 5 1 36 3 1 4 2 5 6 77 5 1 2 3 4 7 68 4 2 1 3 7 6 59 7 5 2 1 3 4 610 7 2 5 4 6 1 311 5 7 6 1 3 2 412 3 5 6 4 7 1 213 4 3 1 2 7 5 614 6 4 7 5 3 1 215 6 3 7 4 5 1 216 7 6 1 3 5 2 417 7 6 3 4 5 1 218 3 1 2 5 4 6 719 3 4 1 2 5 6 720 3 5 2 1 4 6 7Rj 105 70 61 67 97 71 89Table A.5: Ranks R(Xij) and rank sums Rj of prostate V100 given in Ta-ble A.2, for use in the Friedman statistical test and post-hoc analysis.71Appendix A. Data Tables and FiguresPatient A B C D E X Y1 7 2 1 4 3 6 52 7 5 1 6 4 3 23 5 6 1 2 7 3 44 7 2 1 6 5 3 45 5 2 7 4 6 1 36 3 1 4 2 5 6 77 5 1 2 3 4 7 68 4 2 1 3 7 6 59 7 4 2 1 5 3 610 7 1 5 2 6 4 311 7 3 2 1 5 4 612 2 6 3 5 7 1 413 4 3 1 2 7 5 614 6 4 7 2 3 1 515 7 3 6 5 4 1 216 7 5 1 4 6 2 317 7 6 3 5 4 1 218 3 1 2 5 4 6 719 3 4 1 2 5 6 720 3 5 1 2 4 6 7Rj 106 66 52 66 101 75 94Table A.6: Ranks R(Xij) and rank sums Rj of prostate D90 given in Ta-ble A.3, for use in the Friedman statistical test and post-hoc analysis.72Appendix A. Data Tables and Figures01234567Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for BASFigure A.1: Volumetric and dosimetric results for the BAS. Mean manualcontouring values and 95% CIs (blue crosses and error bars) are shown alongwith values derived from model X (red circles) and Y (green squares) con-tours. Overplotted are the dosimetric ?sub-optimal? QA thresholds (dottedlines): V100= 85% and D90= 130 Gy and 180 Gy.73Appendix A. Data Tables and Figures24681012Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for MASFigure A.2: Volumetric and dosimetric results for the MAS. Refer to Fig-ure A.1 caption for a description.74Appendix A. Data Tables and Figures?1012345Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for AASFigure A.3: Volumetric and dosimetric results for the AAS. Refer to Fig-ure A.1 caption for a description.75Appendix A. Data Tables and Figures05101520Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for BLSFigure A.4: Volumetric and dosimetric results for the BLS. Refer to Fig-ure A.1 caption for a description.76Appendix A. Data Tables and Figures51015202530Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for MLSFigure A.5: Volumetric and dosimetric results for the MLS. Refer to Fig-ure A.1 caption for a description.77Appendix A. Data Tables and Figures?2024681012Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for ALSFigure A.6: Volumetric and dosimetric results for the ALS. Refer to Fig-ure A.1 caption for a description.78Appendix A. Data Tables and Figures?202468Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for BPSFigure A.7: Volumetric and dosimetric results for the BPS. Refer to Fig-ure A.1 caption for a description.79Appendix A. Data Tables and Figures02468101214Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for MPSFigure A.8: Volumetric and dosimetric results for the MPS. Refer to Fig-ure A.1 caption for a description.80Appendix A. Data Tables and Figures02468Volume (cc)020406080100V100 (%)  ManualModel XModel Y0 2 4 6 8 10 12 14 16 18 2050100150200250D90 (Gy)PatientDose metrics for APSFigure A.9: Volumetric and dosimetric results for the APS. Refer to Fig-ure A.1 caption for a description.81


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