Prostate Cancer Detection fromMagnetic Resonance Images : AData-driven ApproachbyNandinee Fariah HaqB.Sc., Bangladesh University of Engineering & Technology, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENT OFTHE REQUIREMENTS FOR THE DEGREE OFMASTER OF APPLIED SCIENCEinThe Faculty of Graduate and Postdoctoral Studies(Electrical and Computer Engineering)THE UNIVERSITY OF BRITISH COLUMBIA(Vancouver)August 2014c© Nandinee Fariah Haq 2014AbstractThe combination of Dynamic Contrast Enhanced (DCE) images with Dif-fusion Tensor Images (DTI) has shown great potential in prostate cancerdetection. The parametrization of DCE images to generate cancer mark-ers is traditionally performed based on pharmacokinetic modeling. However,pharmacokinetic models make simplistic assumptions about the tissue per-fusion process, require the knowledge of contrast agent concentration in amajor artery, and the modeling process is sensitive to noise and fitting insta-bilities. In this work, this issue is addressed by extracting features directlyfrom the DCE T1-weighted signal intensities without modeling the physicalperfusion phenomenon.In this work, a novel set of data-driven features are proposed which aregenerated by mapping the DCE T1-weighted signal intensity to its principalcomponent space. The optimal set of components is extracted with sparseregularized regression through a Least Absolute Shrinkage and Selection Op-erator (LASSO) model. It is shown that when the proposed features areused to replace pharmacokinetic parameters, the Area under receiver oper-ating characteristics Curve (AUC) is 0.86, with a support vector machineclassifier trained on the peripheral zone of prostate. When the proposed fea-tures are used within the multiparametric MRI (mpMRI) protocol with theDTI feature, the area under ROC was 0.91 for the peripheral zone classifier,and 0.87 for the whole gland classifier. We showed that in 85.0% cases, thempMRI whole gland classifier detected more than 50% area of the tumor.The proposed features were used to generate cancer likelihood maps forthe prostate gland. These likelihood maps show the likelihood of cancer foriiAbstracteach pixel and hence highlight the regions from where the biopsy samplesshould be taken. The generated cancer likelihood maps have the potentialto be used as a reference in MRI-targeted biopsies to decrease the possibilityof missing clinically significant and potentially aggressive tumors.iiiPrefaceThis thesis is based on the data obtained in 2010-2011 for a multiparametricMRI study titled “Multi-parametric MRI for prostate cancer diagnosis andprognosis” (PI: P. Kozlowski). The study was approved by the Clinical Re-search Ethics Board (CREB) of the University of British Columbia (UBCCREB number : H02-70400).Material from Chapter 2 has been published in the SPIE Medical Imaging2014 under the title “Improved parameter extraction and classification fordynamic contrast enhanced MRI of prostate”. This work is co-authored byPiotr Kozlowski, Edward C. Jones, Silvia D. Chang, S. Larry Goldenberg,and Mehdi Moradi 1.A part of Chapter 3 is published in the Joint Annual Meeting of TheInternational Society for Magnetic Resonance in Medicine under the title“Prostate cancer detection from contrast enhanced T1 time course withoutpharmacokinetic modeling” with the same co-authors 2. Further analysisis published in Computerized Medical Imaging and Graphics under the title“A data-driven approach to prostate cancer detection from dynamic contrastenhanced MRI” 3.1N. F. Haq, P. Kozlowski, E. C. Jones, S. D. Chang, S. L. Goldenberg and M. Moradi.Improved parameter extraction and classification for dynamic contrast enhanced MRI ofprostate. Proc. SPIE 9035, Medical Imaging 2014, pages 903511-903511- 11.2N. F. Haq, P. Kozlowski, E. C. Jones, S. D. Chang, S. L. Goldenberg and M. Moradi.Prostate cancer detection from contrast enhanced T1 time course without pharmacokineticmodeling. Proc. of the Joint Annual Meeting ISMRM, 2014.3N. F. Haq, P. Kozlowski, E. C. Jones, S. D. Chang, S. L. Goldenberg and M.Moradi. A data-driven approach to prostate cancer detection from dynamic contrastenhanced MRI. In Press : Computerized Medical Imaging and Graphics, 2014, doi :10.1016/j.compmedimag.2014.06.017.ivPrefaceThe contribution of author was to develop and evaluate the techniquesproposed in these publications. Prof. Moradi was the primary supervisorand Prof. Kozlowski was the co-supervisor for these works. Dr. Jones, Dr.Chang and Dr. Goldenberg were involved in patient recruitment, surgery,MRI imaging and analysis of the wholemount pathology slides.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . xv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.4 Organization of the thesis . . . . . . . . . . . . . . . . . . . . 62 Data collection and pre-processing . . . . . . . . . . . . . . . 82.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2 Data collection protocols . . . . . . . . . . . . . . . . . . . . 102.2.1 MRI imaging protocol . . . . . . . . . . . . . . . . . . 112.2.2 Pathology data . . . . . . . . . . . . . . . . . . . . . . 132.3 Pharmacokinetic parameters extraction . . . . . . . . . . . . 152.3.1 Semi-automatic arterial input function calculation . . 152.3.2 Pharmacokinetic modeling . . . . . . . . . . . . . . . 16viTable of Contents2.4 Registration . . . . . . . . . . . . . . . . . . . . . . . . . . . 182.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 Data-driven feature extraction and classification from dy-namic contrast enhanced MRI . . . . . . . . . . . . . . . . . . 353.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.2 Data-driven parameterization. . . . . . . . . . . . . . . . . . . 383.2.1 Model-free empirical parameter calculation . . . . . . 393.2.2 Principal component analysis of the T1-weighted in-tensity . . . . . . . . . . . . . . . . . . . . . . . . . . 413.2.3 Feature selection by LASSO . . . . . . . . . . . . . . . 433.3 Classification and cross-validation . . . . . . . . . . . . . . . 443.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534 Multiparametric MRI feature extraction and classification 544.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 544.2 Feature extraction and classification . . . . . . . . . . . . . . 564.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685.2 Discussions and limitations . . . . . . . . . . . . . . . . . . . 695.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . 70Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72viiList of Tables2.1 Clinical data of patients. . . . . . . . . . . . . . . . . . . . . . 113.1 Statistics of the three pharmacokinetic parameters in the for-mat of mean (standard deviation) . . . . . . . . . . . . . . . 473.2 Statistics of the six empirical model-free features in the formatof mean (standard deviation) . . . . . . . . . . . . . . . . . . 483.3 Statistics of the first five PCA features in the format of mean(standard deviation) . . . . . . . . . . . . . . . . . . . . . . . 493.4 Area under receiver operating characteristic curve (AUC), sen-sitivity, specificity and slice-level sensitivity with different fea-ture combinations. Slice-level sensitivity is defined as the per-centage of cases where the classifier can detect more than 50%of the total tumor area. AUC is enlisted in the format of mean(standard deviation). . . . . . . . . . . . . . . . . . . . . . . . 504.1 Statistics of the traditinal multiparametric MRI features inthe format of mean (standard deviation). . . . . . . . . . . . 584.2 Area under receiver operating characteristic curve (AUC), sen-sitivity, specificity and slice-level sensitivity with different fea-ture combinations. Slice-level sensitivity is defined as the per-centage of cases where the classifier can detect more than 50%of the total tumor area. . . . . . . . . . . . . . . . . . . . . . . 59viiiList of Tables4.3 Generated average cancer likelihood values in the format ofmean (standard deviation) with different Gleason scores. Thelikelihood scores were calculated using the peripheral-zone clas-sifier trained on the proposed mpMRI features. . . . . . . . . 61ixList of Figures1.1 Typical T2-weighted MRI of prostate illustrating different zones.The Central and Transitional zones are commonly called Cen-tral Gland on the MR images. . . . . . . . . . . . . . . . . . . 22.1 Prostatectomy sectioning device with walls (LW, lateral walls;SIW, superiorinferior walls) inserted in the base (A), multi-bladed knife (B), the rack (C) used to hold the blades whilethey are cleaned in solution. Reproduced with permission from[16]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2 Changes in the relative signal intensity pattern for two differ-ent pixels with significantly different Ktrans values: the pixelwith higher Ktrans value is from a tumor region and pixel witha lower Ktrans value is from a normal region. The pixels arefrom the peripheral zone of the same patient and from thesame MRI slice. . . . . . . . . . . . . . . . . . . . . . . . . . 182.3 Keypoints extracted by Scale Invariant Feature Transform (SIFT)from (a) T2-weighted MRI, (b) diffusion MRI (b=0 image) and(c) pre injection image from DCE MRI. No matching keypointwas found among them. . . . . . . . . . . . . . . . . . . . . . 202.4 AIF calculation example 1 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform.(c) Generated arterial input function. The blue line is theAIF generated by circular Hough Transform, and the red crossrepresents the AIF generated by manual vessel segmentation. . 24xList of Figures2.5 AIF calculation example 2 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform.(c) Generated arterial input function. The blue line is theAIF generated by circular Hough Transform, and the red crossrepresents the AIF generated by manual vessel segmentation. . 252.6 AIF calculation example 3 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform.(c) Generated arterial input function. The blue line is theAIF generated by circular Hough Transform, and the red crossrepresents the AIF generated by manual vessel segmentation. . 262.7 AIF calculation example 4 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform.(c) Generated arterial input function. The blue line is theAIF generated by circular Hough Transform, and the red crossrepresents the AIF generated by manual vessel segmentation. . 272.8 Dice Similarity Coefficient (DSC) values between pathologyand T2-weighted images. The red and blue bars representsDSC before and after registration respectively. . . . . . . . . 282.9 Dice Similarity Coefficient (DSC) values between T2-weightedand DCE images. The red and blue bars represents DSC be-fore and after registration respectively. . . . . . . . . . . . . . 282.10 Dice Similarity Coefficient (DSC) values between T2-weightedand DTI images. The red and blue bars represents DSC beforeand after registration respectively. . . . . . . . . . . . . . . . 292.11 Registration example 1: (a) Peripheral zone tumor of GleasonGrade (3+4+5) from pathology slide is mapped to the corre-sponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice.The green contour represents the boundary of the prostategland and the red contour is the mapped tumor. . . . . . . . . 30xiList of Figures2.12 Registration example 2: (a) Peripheral zone tumor of GleasonGrade (3+3) from pathology slide is mapped to the corre-sponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice.The green contour represents the boundary of the prostategland and the red contour is the mapped tumor. . . . . . . . . 312.13 Registration example 3: (a) Peripheral zone tumor of GleasonGrade (4+3) from pathology slide is mapped to the corre-sponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice.The green contour represents the boundary of the prostategland and the red contour is the mapped tumor. . . . . . . . . 322.14 Registration example 4: (a) Peripheral zone tumor of GleasonGrade (3+4+5) from pathology slide is mapped to the corre-sponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice.The green contour represents the boundary of the prostategland and the red contour is the mapped tumor. . . . . . . . . 333.1 Kinetic curves of Gadolinium concentration versus time fortumor (red) and normal (blue) regions generated by averagingintensities over all regions of interest. . . . . . . . . . . . . . . 393.2 Changes in the relative signal intensity over time for a regionof interest and the illustration of the empirical model-free pa-rameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403.3 Mean Squared Error (MSE) against the regularization param-eter, λ. λMSE is the λ value for the minimum MSE and λMSE1is the largest λ value for which the MSE is one standard errorof the minimum. . . . . . . . . . . . . . . . . . . . . . . . . . 44xiiList of Figures3.4 (a) Peripheral zone tumor marked in pathology slide. (b)Pathology slide registered to corresponding DCE image. (c)The generated cancer likelihood map superimposed on theDCE T1-weighted image. Note that the classifier is only trainedon the peripheral zone. The classifier is trained on all othercases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513.5 (a) Peripheral zone tumor marked in pathology slide. (b)Pathology slide registered to corresponding DCE image. (c)The generated cancer likelihood map superimposed on theDCE T1-weighted image. Note that the classifier is only trainedon the peripheral zone. The classifier is trained on all othercases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513.6 (a) Peripheral zone tumor marked in pathology slide. (b)Pathology slide registered to corresponding DCE image. (c)The generated cancer likelihood map superimposed on theDCE T1-weighted image. Note that the classifier is only trainedon the peripheral zone. The classifier is trained on all othercases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.7 (a) Peripheral zone tumor marked in pathology slide. (b)Pathology slide registered to corresponding DCE image. (c)The generated cancer likelihood map superimposed on theDCE T1-weighted image. Note that the classifier is only trainedon the peripheral zone. The classifier is trained on all othercases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 524.1 (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the correspondingT2-weighted image. (c) The generated cancer likelihood mapsuperimposed on the T2-weighted image. Note that the clas-sifier is only trained on the peripheral zone. The classifier istrained on all other cases. . . . . . . . . . . . . . . . . . . . . 62xiiiList of Figures4.2 (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the correspondingT2-weighted image. (c) The generated cancer likelihood mapsuperimposed on the T2-weighted image. Note that the clas-sifier is only trained on the peripheral zone. The classifier istrained on all other cases. . . . . . . . . . . . . . . . . . . . . 624.3 (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the correspondingT2-weighted image. (c) The generated cancer likelihood mapsuperimposed on the T2-weighted image. Note that the clas-sifier is only trained on the peripheral zone. The classifier istrained on all other cases. . . . . . . . . . . . . . . . . . . . . 634.4 (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Corresponding DCE image. (c) Cancer likelihoodmap generated using the classifier trained on data-driven DCEfeatures only. (d) Corresponding T2-weighted image. (e) Can-cer likelihood map generated using the classifier trained on theproposed mpMRI features, registered to T2-weighted image. . 644.5 (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Corresponding DCE image. (c) Cancer likelihoodmap generated using the classifier trained on data-driven DCEfeatures only. (d) Corresponding T2-weighted image. (e) Can-cer likelihood map generated using the classifier trained on theproposed mpMRI features, registered to T2-weighted image. . 65xivAcknowledgmentsI would like to thank my supervisors Mehdi Moradi and Piotr Kozlowski fortheir continuous support and advice. For the past two years, their guidanceand enthusiasm kept me motivated. Without their help and suggestions thisproject would not have been possible.I would also like to thank my friends and colleagues at the Robotics andControl Lab for their help and suggestions, both academic and personal. Ithas been difficult to stay so far away from my family and friends, but thefriendly environment in the lab made my life easier.Finally, I would like to thank my parents and my little sister, for beingthe most supportive family one could possibly have. I am grateful for all theencouragement and support they provided throughout my life.xvChapter 1Introduction1.1 MotivationProstate is an exocrine gland of the male reproductive system. It is adoughnut-shaped gland about the size of a walnut. It surrounds the superiorportion of the urethra below the urinary bladder. It secrets an alkaline fluidthat makes upto 13-33% of the volume of the semen. The gland is divided intofour glandular zones, namely- peripheral zone, central zone, transitional zoneand anterior fibro-muscular stroma. Figure. 1.1 shows a typical MagneticResonance Image (MRI) of prostate with the zones outlined. The PeripheralZone (PZ) makes upto 70% of the total gland volume and almost 70-80% ofprostatic cancers originate from this portion of the gland.Prostate cancer is the second most frequently diagnosed cancer and sixthleading cause of cancer-related death among males worldwide [39]. In Canadaand other economically developed countries, it is the most frequently diag-nosed cancer in males. According to the Canadian Cancer Society, in theyear 2014, 24% of all the new cancer cases in men will be prostate cancerand 10% of all cancer deaths will be due to prostate cancer [8]. Prostatecancer is also the most common cancer in European males, that accounts for22.8% of the total cancer cases [23], and its rate is increasing in Asian andEastern European countries as well [38]. The risk of prostate cancer increaseswith age and is diagnosed mostly at the age of 65 years or older.Over 95% of the prostate cancers are adenocarcinoma- where the canceroriginates within the gland cells of the prostate. Most adenocarcinomas arefound in the peripheral zone of the prostate. Mortality and morbidity from11.1. MotivationFibromascular StromaPeripheral ZoneCentral GlandFigure 1.1: Typical T2-weighted MRI of prostate illustrating different zones.The Central and Transitional zones are commonly called Central Gland onthe MR images.prostate cancer is often due to the metastasis of cancer to bones, whichhappens only in a small percentage of cases. In fact, prostate cancer, ifdetected at an early stage, can be cured. The majority of prostate cancersgrow slowly and respond well to treatments. In Canada, the 5-year survivalrate of prostate cancer patients is 96% compared to the healthy people whodo not have cancer [8]. Therefore an early stage diagnosis and staging of theprostate cancer can play an important role in selecting the optimal course oftreatment and hence reduce the mortality rate.The process of prostate cancer screening is controversial. Digital RectalExamination (DRE) is one way of prostate cancer screening, in which theexaminer checks for any abnormalities in the size, shape and texture of theprostate under the clinical fact that cancerous tissue differs in stiffness thannormal tissues. However, DRE is limited to large peripheral lesions. Anotherway of prostate cancer screening is the measurement of the blood level ofProstate Specific Antigen (PSA). PSA test can find prostate abnormalities21.1. Motivationbetter than DRE, and generally PSA levels of 4 ng/ml or higher are takento be associated with high risk of tumor. However, studies have shown thateven patients with lower PSA levels can be at the same risk level [15, 82],and in such cases PSA test can miss the high risk tumors. PSA tests alsoresult in false positives, where elevation of PSA level for other reasons aremisinterpreted as the presence of carcinoma. In such cases, PSA testingcan lead to over-diagnosis and over-treatment, that can potentially result inincontinence, impotence or perioperatory death [94].Since the existing diagnostic techniques cannot adequately determine thestage of prostate cancer, radical prostatectomy is often used in treating thedisease, including instances when the cancer is either pathologically insignifi-cant at one extreme or incurable at the other [83]. This surgery can impact aman’s quality of life by causing urinary incontinence or sexual dysfunctions.Although the techniques associated with prostate cancer imaging and image-guided prostate intervention have improved remarkably over the past decade[34], there is a need for more reliable techniques to confirm cancer location,extent and its stage, prior to deciding on the optimal course of treatmentin a given individual. With reliable imaging techniques, approaches such asfocal therapy and active surveillance of the patients with lower risk tumors,and avoidance of surgery in non-operable cases become possible.Due to the limitations of prostate cancer screening techniques, the mostcommonly used method of prostate cancer diagnosis is a Transrectal Ultra-sound (TRUS) guided biopsy. During biopsy, a radiologist extracts severaltissue samples (6-12) from the prostate through a thin needle. The needle isinserted through the rectum and into the prostate under TRUS. The tissuesamples are then microscopically examined and if cancer is present, the stageand aggressiveness is reported by assigning a Gleason grade. Based on thepattern and growth of the cancer cells, tissue patterns are assigned a numberon the scale of 1 to 5, where higher numbers are given to the aggressive can-cer patterns with less favourable prognosis. The pathologist finds two types31.1. Motivationof growth patterns (primary and secondary patterns) and the summation oftheir Gleason grade is assigned as the final Gleason score or Gleason sum.Although, TRUS provides fine details about the boundaries of the glandand bladder, it has a limited ability to accurately image prostate tissuesand differentiate malignant from benign tissues [31]. Radiologists gener-ally interpret hypoechoic lesions as cancers, while studies have shown thatisoechoic and hyperechoic lesions can be malignant too [17, 36, 78]. As aresult, TRUS-guided biopsy is essentially a blind, systematic biopsy whichsuffers from gross sampling errors. Small cancers may be missed and largeor multifocal tumors may be under-sampled (false negative) [63]. The sensi-tivity of TRUS-guided biopsies are reported to be 40%-60% [18, 63], whichindicates that almost half of the cases tumor samples are missed in TRUS-biopsy. Hence there is a need for diagnostic techniques that can find andstage prostate cancer accurately and non-invasively.Since the mid-1980s, Magnetic Resonance Imaging (MRI) has shown greatpotential in visualization and characterization of prostate anatomy using T2-weighted imaging [6, 33, 72]. The application of endorectal coil [56, 81, 85]and rapid fast-spin echo techniques [58] significantly improved the quality ofMRI images to visualize the internal tissue patterns of the prostate. With itssuperior soft tissue imaging quality, MRI can localize prostate tumors moreaccurately than TRUS [70]. Hence MRI has the potential to improve theperformance of image-targeted biopsies. However, the sensitivity and speci-ficity of the conventional T2-weighted MRI is insufficient for prostate cancerdiagnosis and particularly grading of cancer [2, 25, 54]. As a result, the addi-tion of multiple MR modalities (multiparametric-MRI), such as spectroscopy,diffusion imaging, and perfusion imaging, are under study for better tissuecharacterization [44].Among other MR modalities, Dynamic Contrast Enhanced (DCE) MRIis a very promising methodology [4]. In DCE MRI, a series of T1-weightedMR images are captured after the administration of a contrast agent and41.2. Objectivethe tissue perfusion pattern is estimated by fitting a pharmacokinetic model.The pharmacokinetic parameters extracted from DCE MR images providephysiological information about tissue microvasculature and may be linkedto active cancer growth [7, 66]. These parameters, when used in a super-vised machine learning framework with Diffusion Tensor Imaging (DTI) fea-tures [43, 61], or with T2-weighted and diffusion weighted features [65], haveshown great potential in prostate cancer detection. However, the extractionof the traditional DCE parameters needs information beyond contrast agentconcentration in the prostate and the extracting process is time-consuming,complex and sensitive to noise [20, 75], making it less suitable for clinicaluse. Since the traditional DCE parameters suffer from several limitations,improving the parametrization of DCE MRI is a pressing need.1.2 ObjectiveThe objective of the research in this thesis is to extract data-driven imaging-biomarkers from DCE MR images that can be helpful in prostate cancerdiagnosis. Our approach is to devise a learning agent that can detect cancerdirectly from the DCE T1-weighted intensity without modeling the physi-cal perfusion phenomenon and generate a single parameter map of cancerlikelihood, that can provide the guidance for MRI-targeted biopsies.1.3 ContributionsThe major contributions of this work are:• A semi-automatic approach to vessel extraction and arterial input func-tion calculation is proposed. This approach simplifies the pharmacoki-netic parameter extraction framework.• A method for data-driven parametrization of DCE T1-time course is51.4. Organization of the thesisproposed. This approach is based on principal component analysis ofthe DCE T1-weighted intensities and does not need any pharmacoki-netic modeling. Hence the approach is data-driven and model-free.This approach can simplify the use of DCE data for diagnosis and po-tentially enhance the correlation with pathologic findings.• Cancer likelihood maps are generated for the prostate, which can beused in MRI-targeted biopsies. These maps show cancer likelihood ofeach pixel and hence highlight the region from where the biopsy samplesshould be taken.• A complete framework of prostate cancer detection from multipara-metric MRI is developed. This includes registration of MR images tothe pathology images (used as ground truth), extraction of features,classification and finally generation of single parameter map of cancerlikelihood of the prostate.1.4 Organization of the thesisIn this chapter, the limitations associated with current prostate cancer diag-nostic techniques and the importance of developing new approach to confirmcancer location and extent are discussed. Also we discussed the potentialof MRI in prostate cancer detection, and the significance of exploring newapproaches to parameter extraction. Subsequent chapters are organized asfollows:• Chapter 2 describes the data collection protocol and the pre-processingsteps. It includes the registration of pathology and MR images, andthe pharmacokinetic parameters extraction from DCE MR images bysemi-automatic arterial input function calculation method.• Chapter 3 describes our proposed data-driven parametrization of DCET1-weighted intensities. The technical literature review on machine61.4. Organization of the thesislearning approaches for cancer detection from DCE MR images is in-cluded here.• Chapter 4 includes the performance of the proposed data-driven fea-tures in a multiparametric set-up.• Finally, Chapter 5 includes a summary of the thesis, the limitations ofthe proposed approach and indicates some future works.7Chapter 2Data collection andpre-processing2.1 IntroductionA number of magnetic resonance imaging (MRI) modalities have been inves-tigated for the assessment of prostate cancer. Among other MRI techniques,Dynamic Contrast Enhanced (DCE) MRI has shown promising results indifferentiating normal and cancerous prostate tissues, particularly in a mul-tiparametric framework [43, 61, 65]. Quantitative parameters extracted fromDCE MR images provide physiological information about the tissue microvas-culature and may be linked to active cancer growth [7]. For example, it iswidely reported that one of the pharmacokinetic parameters, volume transferconstant (Ktrans) is higher in peripheral zone tumor tissues than in normalprostate tissues [66]. These quantitative parameters associated with the vas-cular characteristics are extracted by a pharmacokinetic model, usually acompartmental model that delineates contrast agent transfer rate betweenblood plasma and extravascular extracellular space.Pharmacokinetic parameter extraction is an important step in DCE-based cancer detection framework. Another crucial pre-processing step inthe framework of multiparametric MRI-based cancer detection is registra-tion of multi-modality MR images. In this chapter we described :• Pharmacokinetic parameter extraction by semi-automatic arterial in-put function calculation, and82.1. Introduction• Registration of T2-weighted, dynamic contrast enhanced and diffusionMR images to the pathology slides.Pharmacokinetic modeling and Arterial Input Function Estimation:Accurate pharmacokinetic modeling of DCE MRI data requires the knowl-edge of the concentration of the contrast agent in the blood plasma, and thethe concentration is extracted from the the external iliac or femoral arteriesover time, which is known as the Arterial Input Function (AIF). The extrac-tion is generally done by manually segmenting the vessels of interest fromthe MR images. However, this manual approach is susceptible to user biasand partial volume effects.Some automatic methods of vessel pixel identification and AIF extractionare also reported in literature. In Rijpkema et al. an automated vessel pixelselection method was proposed by setting a threshold to the maximum con-trast enhancement value and these pixels were used to generate AIFs [74].In Parker et al. a modified method of vessel pixel isolation is proposed basedon their early enhancement characteristics, and the pixels that show earlyenhancement are assumed as the pixels in the arteries and these were usedto extract AIFs [67]. In Chen et al. the vessel pixels are extracted based onthe peak height and fast uptake characteristics of their voxel intensity curves[11]. However, in all these methods the signal intensity curves are exam-ined for all the pixels of the image, which is time consuming. Therefore anautomatic vessel detection method with reasonable computational time andminimum user interaction would facilitate implementation of an automaticcancer detection framework. In this chapter, we described a semi-automaticalgorithm to contour the cross section of the femoral artery in the DCE im-ages and then to calculate the arterial input function. This approach has thepotential to streamline the use of DCE in cancer detection.Registration :Another important step in the framework of MRI-based cancer detec-92.2. Data collection protocolstion is registration. Ensuring proper alignment of images from different MRmodalities is a pre-requisite to multiparametric MR-based cancer diagnosis.In DCE a series of T1-weighted images is obtained over the span of severalminutes with inevitable patient motion. Another MR modality used in mul-tiparametric MRI-based prostate cancer detection studies is Diffusion TensorImaging (DTI), that maps the diffusion of water molecules in tissues. In dif-fusion imaging the geometric distortion caused by magnetic susceptibility,among other effects, results in non-rigid deformation [48]. Registration ofDCE or diffusion maps to high resolution T2-weighted images and to thehistology images are required for geometric correction and visualization ofthe cancer detection results. In order to align the prostate volume with highprecision, there is a need for a deformable registration methodology to aligna given prostate region across different MR modalities. However, most ofthe studies on multiparametric MR-based cancer detection use qualitativeanalysis based on image intensity where cancer regions are mapped to thecorresponding MR images by a radiologist [14, 28, 45, 46, 62, 66]. In thischapter we described a registration framework to map tumor regions markedin the pathology image to the MRI domain. The employed method uses affinetransformation followed by B-spline registration to map regions of interestfrom the pathology slides to their corresponding T2-weighted, DTI and DCEMR images.2.2 Data collection protocolsThe data used in this work was obtained in 2010-2011 for a multiparametricMRI study (PI: P. Kozlowski). The study was approved by the Clinical Re-search Ethics Board of the University of British Columbia and the patientshad given their written consent before entering the study. These patientswere scheduled for radical prostatectomy and they went through an MRIimaging session before their surgery. The MRI examination was scheduled102.2. Data collection protocols3-48 days prior to their surgery date and the mean time between the MRIsession and the radical prostatectomy was 14 days for these patients. Thepatients recruited for this study had not received any form of therapy beforetheir radical prostatectomy. 21 patients underwent the MRI session out ofwhich 5 patient were excluded from DCE MR imaging due to their aller-gic reaction to the contrast agent. In this work we used MRI data from 16patients who underwent both DTI and DCE MRI sessions. Detailed infor-mation about the patients are given in Table 2.1.Table 2.1: Clinical data of patients.Patient Age PSA Pathology ProstatectomyNumber (years) (ng/mL) stage Gleason score01 66.8 6.5 pT3c 3+402 63.9 2.6 pT2c 3+303 60.9 8.3 pT3c 4+4, 3+4, 3+4+504 52.7 4.9 pT2c 3+305 62.3 7.2 pT3a 3+3, 3+406 67.5 5.3 pT3a 3+3, 3+407 61.3 2.9 pT3a 3+3, 3+408 63.0 5.8 pT2c 3+409 66.3 11.4 pT2a 3+310 69.0 5.4 pT2a 3+311 49.6 6.6 pT3b 3+3, 3+4, 4+312 61.8 7.4 pT3a 3+3, 3+413 56.8 11.0 pT3a 3+4, 3+4+514 64.4 n/a pT2c 3+3, 3+415 67.2 7.0 pT2a 3+416 64.8 4.9 pT2c 3+3, 3+42.2.1 MRI imaging protocolMRI examinations were performed on a 3 Tesla MRI scanner (Achieva,Philips Healthcare, Best, The Netherlands) and the signals were acquired112.2. Data collection protocolswith a combination of an endorectal coil (Medrad, Pittsburgh, PA) and acardiac phased-array coil (Philips Healthcare, Best, The Netherlands). Fastspin-echo T2-weighted images were acquired in the axial and coronal planesusing repetition time (TR) of 1851 ms and an effective echo time (TE) of80 ms with 14 cm field of view (FOV) (284×225 matrix, 3 averages). Eachslice was 4 mm thick and there were no gaps between the slices. 12 axialslices were selected from this sequence and used for DCE MRI scans. T2-weighted images were used to identify the anatomical details of the prostategland to match MR-slices with histology. The majority of the glands weresmaller than 48 mm along the slice selection direction.DCE T1-weighted images were acquired using a three-dimensional T1-weighted spoiled gradient echo-sequence with a field of view of 24 cm (TR/TE= 3.4/1.06 ms, flip angle = 15◦, 256×163 matrix, 2 averages). The contrastagent used here was Gd-DTPA (Magnevist, Berlex Canada) and 0.1 mmol/kgof Gd-DTPA was injected with a motorized power injector within 10 s at therate of 2 mL/s, followed by a 20 mL flush of saline. To calculate the contrastagent concentration in the prostate, at first proton density (PD) images wereacquired (TR/TE = 50/0.95 ms, flip angle = 5◦). Subsequently a series of75 T1-weighted dynamics were acquired, where 3 dynamics were acquiredbefore the injection of the contrast agent, and the remaining 72 dynamicswere acquired during and after the injection of the contrast agent. Thetime resolution was 10.6 s per dynamic and the slice thickness was 4 mm.T1 values were calculated based on PD-weighted and T1-weighted imagesaccording to the procedure described by Parker et al [68]. DCE MRI datawere processed off-line using Matlab (Mathworks, Natick, MA) and Igor Pro(WaveMetrics, Portland, OR).DTI data were acquired using a diffusion weighted single shot echo planarimaging (EPI) sequence with a field of view of 24 cm (TR/TE = 2100/74 ms,slice thickness = 4 mm with no gap, 128 × 115 matrix, 6 noncollinear gra-dient directions, 18 averages, total acquisition time of 8 min, b-value = 0122.2. Data collection protocolsand 600 s/mm2). DTI data were processed off-line to calculate fractionalanisotropy (FA) and average diffusivity (〈D〉) values. Diffusion weightedimages were registered to the non-weighted b=0 image with a mutual in-formation algorithm and eigenvalues of the diffusion tensor were calculated.Average diffusivity and fractional anisotropy maps were generated with theproprietary DTI processing toolbox PRIDE (Philips Healthcare, Best, TheNetherlands).2.2.2 Pathology dataAfter imaging, patients went through the surgery and radical prostatectomyspecimens were acquired. The prostatectomy specimens were dissected andhistopathologically examined in a uniform manner to acquire the whole-mount sections. The external surfaces were inked, seminal vesicles were am-putated, the apical and bladder neck tissue slices were removed and the spec-imens were dissected following a minimum of 24-h fixation in 10% bufferedformalin. A device described in Drew et al [16] was used to cut the prostategland from inferior to superior in serial transverse cuts perpendicular to theposterior capsule, at 4 mm intervals, which allowed reasonably good corre-spondence between the pathology slides and the MR slices.Figure. 2.1 shows different parts of the prostatectomy sectioning device.The multi-bladed cutting device consists of an adjustable box that holds theprostate, a plunger tool that prevents rolling or sliding of prostate, and a knifeassembly that holds the blades at 4 mm intervals. Once the prostate specimenis placed on the base, the specimen is rotated manually to the orientationthat produces the proper slice orientation. The adjustable box was designedto hold the prostate in the correct orientation for matching the MRI and thehistology sections. The walls of the device and the plunger include verticalslots to allow the blades of the knife to pass through it. The slots are at 4 mmintervals and the slots on opposite walls are lined up with each other to ensureparallel cuts. The blades of the knife assembly fit through the vertical slots132.2. Data collection protocolsFigure 2.1: Prostatectomy sectioning device with walls (LW, lateral walls;SIW, superiorinferior walls) inserted in the base (A), multi-bladed knife (B),the rack (C) used to hold the blades while they are cleaned in solution.Reproduced with permission from [16].in the adjustable box and cut simultaneously through the sample. The multi-blade device ensures that the histology sections are parallel to each other andexactly 4 mm apart - requirements that are difficult to achieve with manualcutting using a single blade. In addition, careful positioning of the specimeninside the cutting device results in the orientations of the MRI slices andhistology sections being very close to each other.A Lieka RM2245 whole body rotary microtome was used to cut the wholemount sections and the sections were submitted as intact transverse sectionsmounted on oversize glass slides for hematoxylin and eosin staining. In thepathology slides, the regions of prostatic carcinoma were outlined and as-signed a Gleason score by an anatomic pathologist with over 20 years ofexperience.142.3. Pharmacokinetic parameters extraction2.3 Pharmacokinetic parameters extractionIn order to compare our data-driven approach proposed in Chapter 3 to thetraditional pharmacokinetic parameter based approach to prostate cancer de-tection, we generated pharmacokinetic parameter maps for the patients. TheArterial Input Function (AIF) was calculated by a semi-automatic methodand from the extracted AIFs local population average AIF signal was formed,which was then used to generate pharmacokinetic parameter maps using two-compartmental Kety model.2.3.1 Semi-automatic arterial input functioncalculationWe present a semi-automatic technique of extracting patient-specific AIFthat can generate AIF similar to that calculated by manual segmentation.In this method, operator interaction is limited to selecting a seed point forthe vessel. The algorithm searches for circular structures with a radius of 3to 8 mm within a 40 mm × 40 mm area centered at the seed point using thecircular Hough transform algorithm [3]. Circular Hough transform identifiespossible circular objects according to the equation of a circle:(i− a)2 + (j − b)2 = r2 (2.1)Here, (a, b) is the center and r is the radius of the circle. (i, j) are theedge points of the image of interest extracted from the gradient image. Cir-cular Hough transform creates an accumulator space made up of cells thatinclude all the possible values of circle centers. For each edge pixel (i, j) thealgorithm keep incrementing all such cells that could be the center of thecircle according to Equation 2.1. Then the algorithm searches for the localmaxima in the accumulator space, and finds the circle centers that have thehighest probability of being a circle. With this algorithm, the pixels inside152.3. Pharmacokinetic parameters extractionthe vessel were identified.Assuming minimum partial volume effect, the extracted vascular pixelswere then used to generate the concentration as a function of time using thefollowing equation [68]:C(t, p) =KTRln(Spd(p)− Spre(p)Spd(p)− Spost(t, p))(2.2)where C(t, p) is the contrast agent concentration in time t at pixel p, Spd is thesignal intensity in the proton density image, Spre is the pre-injection signalintensity, Spost is the pixel intensity at time t post-injection, K is a constantthat reflects the contrast agent relaxivity and pulse sequence parametersand TR is the repetition time. From Equation 2.2, the concentration valueswithin 5% of the peak concentrations were selected and averaged to createthe final AIF [67].Patient-specific AIFs were extracted from the external iliac or femoralarteries in the center-slice of the DCE MRI data. The MR images used inthis work were acquired using the 3D technique. This technique minimizesthe blood in-flow effect, as opposed to the 2D multislice technique; and there-fore slice selection has a minimal effect on the arterial input function. Weselected the center-slice to extract AIF as this slice best visualizes the arter-ies. We used local population averaged AIF, called local Gaussian AIF [59],to calculate the pharmacokinetic parameters. Here the AIF is calculatedin the form of two Gaussians plus an exponential function, as proposed byParker [69], but fitted to the average of individual AIFs measured from thepatients in our study. This functional form improves the reproducibility ofthe pharmacokinetic parameters.2.3.2 Pharmacokinetic modelingThe extracted AIF was then used to calculate three pharmacokinetic param-eters by fitting the extended Kety model to the contrast agent concentration162.3. Pharmacokinetic parameters extractioncurves [88]. The two-compartmental Kety model is a standard tracer kinet-ics model, which assumes that the tracer (contrast agent) can reside in twocompartments: blood plasma space and extra-vascular extra-cellular space(EES), under the assumption that the tracer does not cross the cellular mem-branes. Following intravenous injection of a bolus of a contrast agent, theagent passes through the arteries and its concentration can be measured inthe supplying arteries - known as the AIF. Transfer of the contrast agent fromblood plasma to interstitial space is driven by the concentration gradient. Ini-tially, when blood plasma concentration is high, the agent diffuses into EESthrough the capillaries microscopic pores. When the plasma concentrationbecomes lower than the EES concentration, the contrast agent diffuses backto the plasma space. Overall concentration of the agent (plasma + EES) isadditionally affected by the clearance of the agent through kidneys and liver.The rate at which this transfer of contrast agent between plasma and EESoccurs is known as the volume transfer constant, Ktrans.The volume transfer constant (Ktrans) depends on the permeability ofcapillary wall and the flow of blood. In the flow-limited condition (i.e. whenthe vessels are very leaky and the contrast agent leaves the plasma spacevery quickly), Ktrans is equal to the blood plasma flow per unit volume oftissue. In the permeability-limited condition (i.e. when the vessels’ perme-ability is very low), Ktrans is equal to the permeability surface area productper unit volume of tissue. The fractional volumes of the EES and plasmacompartments are known as ve and vp respectively. These three pharma-cokinetic parameters: volume transfer constant, Ktrans, fractional volume ofextravascular extracellular space, ve, and fractional plasma volume vp werecalculated for every pixel of every slice within a region of interest encompass-ing the prostate gland.Figure.2.2 shows the relative signal intensity pattern for two regions, withsignificantly different Ktrans values. The pixel with higher Ktrans value isfrom a tumor region and pixel with a lower Ktrans value is from a normal172.4. Registration0 2 4 6 8 10 12 14−0.500.511.522.533.5Time (minutes)Relative signal intensity Ktrans = 0.08Ktrans = 0.02Figure 2.2: Changes in the relative signal intensity pattern for two differentpixels with significantly different Ktrans values: the pixel with higher Ktransvalue is from a tumor region and pixel with a lower Ktrans value is from anormal region. The pixels are from the peripheral zone of the same patientand from the same MRI slice.region. As can be seen from the figure, region with high Ktrans exhibitshigher and faster intake and faster washout rate than the region with lowKtrans. It should be noted that the pharmacokinetic parameters measurephysical quantities such as the transfer rate. The transfer rate also drives thechanges in the image intensity. As such, not surprisingly, as Figure.2.2 shows,higher Ktrans values can be attributed to a fast increase in the intensity afterinjection.2.4 RegistrationOur tissue cutting protocol and device ensured reasonably good correspon-dence between pathology and MR slides. Hence a 2D registration frameworkcould be used to align MR and pathology slides. However, due to shrink-age of the tissue after fixation, deformable registration was required to mapthe histology findings to the MR images. Registration of DCE images to182.4. RegistrationT2-weighted images were also needed to compensate for the motion of theprostate during the prolonged DCE MRI scan. Moreover, diffusion MR im-ages were required to register to T2-weighted MRI to compensate for thegeometric distortion caused by magnetic susceptibility. One approach toregister different images is landmark-based registration, where anatomical orfiducial landmarks are used to guide the registration procedure. However,we do not have implanted fiducial landmarks and we could not find stableanatomical landmarks in the pathology, T2-weighted, DTI and DCE MRimages. Hence landmark-based registration was not carried out.An alternative approach to register images without anatomical and artifi-cial landmarks is to extract scale space keypoints by Scale Invariant FeatureTransform (SIFT) and use these keypoints to guide the registration algorithm[60]. SIFT extracts the keypoints from the image by extracting local featuresthat are scale and orientation invariant and robust to illumination changes[52]. A descriptor vector is assigned to describe each keypoint, and the de-scriptor vector is created by computing the gradient at each pixel aroundthe keypoint. The extracted keypoints can be used with landmark-basedregistration algorithm where the matching point for each keypoint in themoving image is found by identifying its nearest neighbor from the set ofthe keypoints extracted from the fixed image. However, in this work, SIFTtransform failed to generate matching scale-invariant features between T2-weighted and DTI or DCE MR images. Figure 2.3 shows keypoints extractedfrom corresponding T2-weighted, DTI and DCE MR images. The keypointsare visualized as vectors where the origin, length and direction of the vectorrepresents location, scale and orientation respectively. As can be seen fromFigure 2.3, no matching keypoint between the images was identified.One solution for the registration problem is to use intensity-based reg-istration that does not need anatomical or artificial landmarks. The regis-tration method used in this work was intensity-based affine transformationfollowed by deformable B-spline registration as proposed in [77]. In this al-192.4. Registration(a) (b) (c)Figure 2.3: Keypoints extracted by Scale Invariant Feature Transform (SIFT)from (a) T2-weighted MRI, (b) diffusion MRI (b=0 image) and (c) pre injec-tion image from DCE MRI. No matching keypoint was found among them.gorithm, the transformation T that maps the moving image coordinates tothe fixed image domain is a combination of a global (affine) tranformation,Taffine and a local (B-spline) transformation, TB−spline. For each pixel (i, j),the affine transformation is defined as:Taffine(i, j) = Θ1x + Θ2 (2.3)where x = [i j]T . Θ1 accounts for the rotation, scale and shear of the movingimage pixel and Θ2 accounts for the translation of the pixel.To account for the the deformation in prostate B-spline transformation isemployed. In B-spline transformation, a number of control points are definedat equal intervals, σ and the image is deformed by manipulating this meshof control points. If xˆ denotes the mesh of ni × nj control points xˆi′,j′ , theB-spline transformation can be written as [77]:TB−spline(i, j) =3∑m=03∑n=0Θm(p)Θn(q)xˆi′+m,j′+n (2.4)where i′ = bi/nic − 1, j′ = bj/njc − 1, p = i/ni − bi/nic, q = j/nj − bj/njcand the B-spline basis functions can be written as [49, 50]:202.4. RegistrationΘ0(p) =(1− p)36Θ1(p) =3p3 − 6p2 + 46Θ2(p) =−3p3 + 3p2 + 3p+ 16Θ3(p) =p36Mutual information was used as the similarity measure for the registrationalgorithm. If Ia and Im are fixed and moving image respectively, then themutual information between Ia and transformed moving image Ib = Im ◦ Tis formulated as:MI(Ia, Ib) =∑a∈La∑b∈Lbp(a, b) log2p(a, b)pA(a)pB(b)(2.5)Here La and Lb are sets of regularly spaced intensity bin centers, pA and pBare the marginal discrete probability, and p is the discrete joint probabilityestimated using B-spline Parzen window described in [57].Adaptive Stochastic Gradient Descent (ASGD) algorithm was used as theoptimizer to guide the registration process. In a registration framework withmutual information as the similarity measure, the optimization problem isto solve for the vector of transformation parameters, µˆ such that:µˆ = arg minµMI(Ia, Ib) (2.6)ASGD algorithm solves this minimization problem iteratively by search-ing in the direction of negative gradient of the cost function (in this casemutual information). ASGD algorithm is formulated as:212.4. Registrationµk+1 = µk − ζ(tk)∆ˆk (2.7)where, ζ(tk) =α(tk + A)β(2.8)and tk+1 = max(tk + f(−∆ˆTk ∆ˆk−1), 0)(2.9)Here ∆ˆk is the approximated derivative of the cost function, MI, µk isthe approximated transformation parameter at kth iteration and f denotes asigmoid function. A, α and β are user defined constants. Detailed descriptionof the algorithm can be found in [41].Using this registration framework, wholemount pathology slides were reg-istered to their corresponding T2-weighted MRI, and T2-weighted MRI sliceswere registered to the corresponding DCE and DTI slices. To register theimages, the prostate regions were manually segmented in pathology, T2-weighted, DCE and DTI slices, and registration was applied on the segmentedregions only. The segmentation of prostate ensured that the registration al-gorithm only focuses on the prostate region and disregards the deformationsin the muscular areas. After registration, the computed transformations wereapplied to map the tumor regions from the pathology slides to the correspond-ing T2-weighted and to the DTI and DCE-MR images. For registration, weused elastix, which is an open source software for image registration [42].We used Dice Similarity Coefficient (DSC) to validate the registration.Dice Similarity Coefficient between two images, Ia and Ib is defined as follows:DSC(Ia, Ib) =2× total number of overlapping pixelstotal pixels in Ia + total pixels in Ib(2.10)As discussed earlier, we did not have implanted fiducial landmarks andwe could not find stable anatomical landmarks in the pathology and MRimages, hence target registration error was not measured.222.5. Results2.5 ResultsEvaluation of the AIF extraction technique:Figures 2.4 to 2.7 show four typical results from our vessel contouringand AIF extraction method. In each case, the top two images show the DCEimage and the extracted arteries using circular Hough transform. The ploton the second row shows the generated arterial input function. The blue linerepresents the AIF generated by the circular Hough transform and the redone represents the AIF generated by manually selecting a region of interestencompassing the vessels. As can be seen, there is no significant differencein AIF generated by these two methods. Figure 2.6 is the only case wherethere were slight differences in concentration values between the proposedvessel extraction method and manual segmentation method, and the meanabsolute error for this patient was 0.08. The difference in AIFs may havebeen resulted due to the fact that in this case the femoral blood vessels werein close proximity and the the manual method may have included some brightpixels from the neighbouring vessel.Evaluation of the registration method:The DSC values are calculated over a window encompassing the prostate.The average DSC value between the pathology and T2-weighted images afterregistration was 0.93±0.02. The average DSC between T2-weighted andregistered DCE image was 0.95±0.03. The same DSC was recorded for T2-weighted to DTI registration. Figures 2.8, 2.9 and 2.10 show the averageDSC values for 16 patients before and after registration. Figures 2.11 to2.14 show four cases where the tumor regions from the pathology slides aremapped to the corresponding T2-weighted, DTI and DCE MRI slices usingaffine transformation followed by B-spline registration algorithm.232.5. Results(a) (b)0 2 4 6 8 10 12 1402468Time (minutes)Gd−DTP [nMol] Semi−automaticManual(c)Figure 2.4: AIF calculation example 1 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform. (c) Generatedarterial input function. The blue line is the AIF generated by circular HoughTransform, and the red cross represents the AIF generated by manual vesselsegmentation.242.5. Results(a) (b)0 2 4 6 8 10 12 140246810Time (minutes)Gd−DTP [nMol] Semi−automaticManual(c)Figure 2.5: AIF calculation example 2 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform. (c) Generatedarterial input function. The blue line is the AIF generated by circular HoughTransform, and the red cross represents the AIF generated by manual vesselsegmentation.252.5. Results(a) (b)0 2 4 6 8 10 12 1402468Time (minutes)Gd−DTP [nMol] Semi−automaticManual(c)Figure 2.6: AIF calculation example 3 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform. (c) Generatedarterial input function. The blue line is the AIF generated by circular HoughTransform, and the red cross represents the AIF generated by manual vesselsegmentation.262.5. Results(a) (b)0 2 4 6 8 10 12 1402468Time (minutes)Gd−DTP [nMol] Semi−automaticManual(c)Figure 2.7: AIF calculation example 4 : (a) Dynamic contrast enhancedimage. (b) Extracted arteries using circular Hough transform. (c) Generatedarterial input function. The blue line is the AIF generated by circular HoughTransform, and the red cross represents the AIF generated by manual vesselsegmentation.272.5. Results1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 160.50.60.70.80.91PatientsDice Similarity Coefficient After RegistrationBefore RegistrationFigure 2.8: Dice Similarity Coefficient (DSC) values between pathology andT2-weighted images. The red and blue bars represents DSC before and afterregistration respectively.1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 160.50.60.70.80.91PatientsDice Similarity Coefficient After RegistrationBefore RegistrationFigure 2.9: Dice Similarity Coefficient (DSC) values between T2-weightedand DCE images. The red and blue bars represents DSC before and afterregistration respectively.282.6. Summary1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 160.50.60.70.80.91PatientsDice Similarity Coefficient After RegistrationBefore RegistrationFigure 2.10: Dice Similarity Coefficient (DSC) values between T2-weightedand DTI images. The red and blue bars represents DSC before and afterregistration respectively.2.6 SummaryIn this chapter, we described a image pre-processing streamline for multi-parametric MRI-based cancer detection. This includes efficient extractionof arterial input function and pharmacokinetic parameters extraction fromDCE MRI, and registration of multi-modality MR images to pathology slides.We described a semi-automatic AIF extraction technique from DCE MRimages. The framework is semi-automatic and includes vessel extractionby circular Hough transform to calculate arterial input function and usesextended Kety model to generate the pharmacokinetic parameter maps. Weshowed that in all but one case our proposed vessel extraction techniquegenerated AIFs identical to those generated by segmenting vessels manually.We also described a registration framework for multiparametric MRIstudies. By using affine and deformable B-spline transformation, we reg-istered the wholemount pathology slides to their corresponding T2-weightedimages, and T2-weighed images to the corresponding dynamic contrast en-hanced and diffusion MR images. One limitation of this registration method292.6. Summary(a) (b)(c) (d)Figure 2.11: Registration example 1: (a) Peripheral zone tumor of GleasonGrade (3+4+5) from pathology slide is mapped to the corresponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice. The green contour represents theboundary of the prostate gland and the red contour is the mapped tumor.302.6. Summary3+3(a) (b)(c) (d)Figure 2.12: Registration example 2: (a) Peripheral zone tumor of GleasonGrade (3+3) from pathology slide is mapped to the corresponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice. The green contour represents theboundary of the prostate gland and the red contour is the mapped tumor.312.6. Summary(a) (b)(c) (d)Figure 2.13: Registration example 3: (a) Peripheral zone tumor of GleasonGrade (4+3) from pathology slide is mapped to the corresponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice. The green contour represents theboundary of the prostate gland and the red contour is the mapped tumor.322.6. Summary(a) (b)(c) (d)Figure 2.14: Registration example 4: (a) Peripheral zone tumor of GleasonGrade (3+4+5) from pathology slide is mapped to the corresponding (b) T2-weighted, (c) DTI and (d) DCE-MRI slice. The green contour represents theboundary of the prostate gland and the red contour is the mapped tumor.332.6. Summaryis that it deforms the whole prostate gland simultaneously. In some cases thelevel of deformation present in the central gland might be different than theperipheral zone, and in those cases the registration algorithm can fail to reg-ister the gland properly. One possible solution could be to register the centralgland and peripheral zone separately, which can be a future improvement tothis method.34Chapter 3Data-driven feature extractionand classification from dynamiccontrast enhanced MRI3.1 IntroductionDynamic Contrast Enhanced (DCE) MRI has shown great potential in prostatecancer diagnosis and staging. In DCE MRI, a gadolinium based contrastagent is administered intraveneously, and a series of T1-weighted imagesare acquired before, during and after the injection of the contrast agent.The temporal changes in the signal intensity captured with this dynamicacquisition scheme provide useful information about the underlying tissuemicrovasculature. Quantitative parameters extracted from the DCE MRIsignal intensities have the potential to provide physiological information re-lated to the vascular characteristics of the tissue and may be linked to activecancer growth [7, 66]. These parameters are extracted by fitting a tracer-kinetic model, generally a compartmental pharmacokinetic model to the con-trast agent concentration curves [53, 87]. Most commonly used pharmacoki-netic parameters are volume transfer constant (Ktrans), fractional volumeof extravascular extracellular space (ve) and fractional plasma volume(vp).These quantitative parameters have been reported to improve the accuracyof prostate cancer localization compared to T2-weighted MRI [62, 80]. Thepharmacokinetic parameters, when used in a supervised machine learningframework with diffusion tensor features [43, 61], or with T2-weighted and353.1. Introductiondiffusion weighted features [65], have shown great potential in prostate cancerdetection.The quantitative pharmacokinetic parameters are extracted by fitting thesequence of contrast agent concentration values to one of the several phar-macokinetic models that characterize the perfusion pattern. Accurate phar-macokinetic modeling of DCE MRI data also requires the knowledge of theconcentration of the contrast agent in the external iliac or femoral arteriesover time, known as the AIF. The extraction of AIF requires the segmenta-tion of the cross section of the artery, which is a subjective process. Besides,the modeling process is sensitive to noise and fitting instabilities [75], andthe fitting process is complex and difficult for clinical use [20]. Furthermore,most pharmacokinetic models make simplistic assumptions about the perfu-sion process in tissue. This issue has resulted in limited correlation betweenindividual DCE parameters and the stage of the disease [19, 32], despite thestrong evidence of correlation between increase in microvasculature and thepathologic stage of cancer [79, 89]. Since pharmacokinetic modeling is not theideal solution for characterizing the perfusion process based on DCE data,there is a need for pharmacokinetic model-free framework for the analysis ofthe DCE T1-weighted intensities.An alternative approach of extracting useful information from DCE MRIis to analyse the pattern of dynamic T1-weighted intensity curves by extract-ing empirical parameters. Several empirical parameters are reported to havethe potential to be used in prostate cancer diagnosis. Linely et al. reporteda shorter time to reach maximum enhacement in prostate cancer lesions thannormal peripheral zone [51]. Several studies reported early wash-in and fasterwash-out rates of contrast agent in prostatic carcinoma [40, 73, 76]. Padhaniet al. showed significant differences in signal intensity variables, i.e. meangradient and maximum signal intensity [66]. Zelhof et al. evaluated theperformance of model-free empirical features (the maximum enhancementindex, time to reach maximum enhancement and area under signal intensity363.1. Introductioncurve), and reported an accuracy of 90% with logistic regression analysisusing wholemount radical prostatectomy specimens as reference [95].Aside from estimating empirical parameters from the signal intensitycurves, several studies focused on extracting features by mapping the T1-weighted intensities to a low dimensional space. Principal Component Anal-ysis (PCA) and independent component analysis has been investigated foranalysing DCE MRI intensity patterns. Eyal et al. showed that the sec-ond and third eigenvectors correlates with the histopathologic findings andshowed that independent component analysis is less informative as an imag-ing biomarker for prostate cancer diagnosis [21]. The study in [22] showedthat, for intensity scaled DCE MRI, the second principal component provideuseful information related to prostate cancer diagnosis when adjusted by ro-tation to reach congruence with wash-in and wash-out kinetic parameters.In this chapter, we further investigate this approach to develop a frameworkfor MRI-targeted prostate biopsies. Our work combines the principal compo-nent analysis with a new feature selection approach and uses classification toproduce cancer likelihood maps. This approach can simplify the use of DCEdata for diagnosis and potentially enhance the correlation with pathologicfindings.Our approach is to devise a learning agent that can detect cancer directlyfrom the T1-weighted intensity without modeling the physical perfusion phe-nomenon. We investigated two different sets of parameters in terms of theircapabilities to classify prostate cancers. The first set of parameters consistsof six empirical model-free parameters that describe the shape of the kineticcurves of DCE T1-weighted signal intensity. The second set of parametersis generated by reducing the dimensionality of the T1-weighted intensitiesusing Principal Component Analysis (PCA). An important question in theuse of PCA for feature reduction is the determination of the optimal numberof components that will be included in the model. It is also not guaran-teed that only the components that can explain most of the variance of the373.2. Data-driven parameterization. . .data provide the highest classification accuracy. So we have devised a newsolution of extracting the optimal set of features from PCA. To reduce thedimensionality in the PCA space, we used a Least Absolute Shrinkage andSelection Operator (LASSO) model, along with cross-validation to determinethe number of components. The resulting parameters were used with Sup-port Vector Machine (SVM) classification. These parameters were generateddirectly from the DCE MRI signal intensities and therefore do not make anyassumption about the perfusion pattern.3.2 Data-driven parameterization of DCET1-weighted time courseThe normal and cancerous regions were mapped from the pathology slidesto the corresponding DCE slides by affine transformation followed by B-spline registration as described in Section 2.4. The mapped regions werethen used to extract features. To extract features, these mapped regionswere divided into square regions of interest (ROI) of size 4.7 mm × 4.7 mmin the DCE image. The intensity values within each ROI were averaged andan average T1-weighted signal was formed. Thereby, each ROI was used toform one T1-weighted signal and the model-free features were extracted fromthis average T1-weighted signal. The true label (class) of the feature vectorwas determined from the pathology image. Figure 3.1 shows the kineticcurves of contrast agent concentration in the prostate. These curves aregenerated by averaging over all the ROIs of the tumor and normal regions.The difference in the pattern of perfusion is evident: the contrast agentconcentration increases faster and stronger within the tumor.We extracted two sets of model-free features: empirical features that de-scribe the shape of the DCE T1-weighted signal intensity curve, and thefeatures based on principal component analysis of the signal intensities. Fur-thermore, a feature selection algorithm was applied on the second set of383.2. Data-driven parameterization. . .0 2 4 6 8 10 12−0.500.511.52Time (minutes)Gd−DTP concentration [nMol] TumorNormalFigure 3.1: Kinetic curves of Gadolinium concentration versus time for tumor(red) and normal (blue) regions generated by averaging intensities over allregions of interest.features to identify the optimal set of significant features.3.2.1 Model-free empirical parameter calculationRadiologic evaluation of the DCE time course is typically performed by ex-amining the kinetics of the concentration time course [31]. As part of ourdata-driven approach to parameterization of the DCE data, we extracted sixmodel-free empirical features from the DCE T1-weighted signal intensities.These features were extracted from the relative signal intensity calculated bythe following equation:irel(t) =i(t)− ipreipre(3.1)where ipre is the average pre-injection signal intensity over an ROI and i(t)is the average signal intensity at time t. Figure 3.2 shows the change in therelative intensity over time for a single ROI. The features extracted from therelative signal intensity, irel(t) are:393.2. Data-driven parameterization. . .Figure 3.2: Changes in the relative signal intensity over time for a region ofinterest and the illustration of the empirical model-free parameters.Maximum signal intensity (Imax): defined as the peak relative signalintensity in DCE time course.Time-to-peak (tpeak): defined as the time to reach the maximum signalintensity.Onset time (tonset): defined as the time to reach 10% of the maximumsignal intensity.Initial gradient (∆initial): defined as the mean gradient from the timeof injection to the time when the intensity is 10% of the maximum signalintensity.Mean gradient (∆mean): defined as the mean gradient calculated be-tween the time points when signal intensity goes from 10% to 90% of themaximum intensity.Washout gradient (∆washout): defined as the average rate of changein relative signal intensity from the maximum intensity to the end of thescanning period.403.2. Data-driven parameterization. . .3.2.2 Principal component analysis of theT1-weighted intensityAn alternative approach to data-driven characterization of the DCE T1-weighted intensities is to use a dimensionality reduction method to convertthe time series of normalized contrast enhanced T1-weighted intensities to anoptimally sized vector of features. We used Principal Component Analysis(PCA) for this purpose. PCA uses an orthogonal transformation to map ahigh dimensional data of possibly correlated variables into a low dimensionalrepresentation with linearly uncorrelated variables, called principal compo-nents. The first principal component has the largest variance, or it inheritsthe maximum variability of the data. Each successive component in turnaccounts for the highest variance possible under the constraint that it isorthogonal to the preceding components [1].If we have p number of dynamic post-injection images, then each regionof interest gives rise to a p-dimensional observation vector representing thedynamic image intensities over p-time points: i = {irel(1), irel(2), ..., irel(p)},where irel(p) is the relative signal intensity at time point p. If the high-dimensional training data-set consists of N observations, it can be written asa zero-empirical mean data-set: X = {x1; ...; xN} , where each observation,xi is a p-dimensional row vector constructed by subtracting the sample meanfrom ii. PCA projects the data to a new p-dimensional orthogonal coordinatesystem such that:I = XW (3.2)where, I for the N × 72 dimensional matrix obtained by the PCA transfor-mation of the 72-dimensional T1-weighted intensities for all the N trainingROIs and W is the matrix of coefficients that projects X into its principalcomponent space. PCA tries to find these row factor scores, obtained as alinear combinations of columns of X such that these factor scores explain413.2. Data-driven parameterization. . .maximum possible variance of X and the sets of factor scores are pairwiseorthogonal. The constraint imposed here is that the sum of squares of thecoefficients of the linear combination is equal to unity. Therefore the con-straints on Equation 3.2 are [1]:1. I is an orthogonal matrix, or IT I is a diagonal matrix, whereIT I = WTXTXW (3.3)2. W is an orthonormal matrix, orWTW = Ip (3.4)The solution of this problem can be obtained with Lagrangian multipliers,by defining the following equations and solving for W that maximizes ` [30].` = trace{IT I−∆(WTW − Ip)} (3.5)The operator trace gives sum of the diagonal elements of a square matrixand ∆ is a diagonal matrix of Lagrangian multipliers. Differentiating Eq.3.5 with respect to W and setting it to zero gives us:XTX = W∆WT (3.6)Since ∆ is a diagonal matrix, this is an eigen-decomposition problem,where ∆ is the matrix of eigenvalues of XTX ordered from largest to small-est eigenvalue and W is the matrix of its associated eigenvectors. Also,combining Eqs. 3.6 and 3.3 we can write:IT I = WTXTXW = ∆ (3.7)Hence the variance of the factor scores are essentially equal to the eigen-values. Since the sum of the eigenvalues is the trace of XTX, this concludesthat the first factor score has the highest variance [1]. For a p-dimensional423.2. Data-driven parameterization. . .data-set, PCA analysis gives p number of eigenvectors and projects the datainto a p-dimensional principal component space.3.2.3 Feature selection by LASSOThe optimal set of principal components is extracted with sparse regular-ized regression through Least Absolute Shrinkage and Selection Operator(LASSO). The 72-dimensional principal components extracted from the DCET1-weighted intensities were ranked in terms of class separability using LASSO[86]. The L1-norm regression problem in our work was formulated as:βˆ = arg minβ(12N∑j=1(yj − β0 −p∑k=1Ijkβk)2 + λp∑k=1|βk|) (3.8)where Ijk are elements of I, yj is the label (0 or 1) for the jth ROI, λ is theregularization parameter and the vector β includes the LASSO coefficients.This L1-norm regression forces some coefficients to zero. The 72 features areranked according to the magnitude of their corresponding βk. To solve theregression problem, cylindrical coordinate descent algorithm was used [24]and the value of λ was determined by 10-fold cross validation on the en-tire dataset targeted to minimize the mean squared error. Figure 3.3 showsthe mean squared error against the values of λ. Among 72 PCA features,44 components were identified as the significant features by LASSO withnon-zero coefficients. Leave-one-patient-out cross-validation with forwardsearch showed that the maximum Area Under receiver operating character-istics Curve (AUC) was obtained when the 22 top-ranked PCA features wereused together to train and test the classifier.433.3. Classification and cross-validation10−410−310−210−10.30.40.50.60.70.80.911.1LambdaMean Squared ErrorCross−validated MSE of Lasso fitλMSE1 λMSEFigure 3.3: Mean Squared Error (MSE) against the regularization parameter,λ. λMSE is the λ value for the minimum MSE and λMSE1 is the largest λvalue for which the MSE is one standard error of the minimum.3.3 Classification and cross-validationIn this work, support vector machine (SVM) classifier was used to classifycancer and normal tissues with different feature combinations. The use ofSVM enables us to perform classification without the need for estimating theprobability distributions of the features, as needed in Bayesian approaches.The SVM solution is also a convex optimization problem, which reduces therisk of local minima.SVM classifier constructs a maximum-margin hyperplane or a set of hy-perplanes in a high-dimensional space to separate the input data into differ-ent classes. Typically, a kernel function is used with SVM to map the datainto a high-dimensional space. The optimization process that trains SVM isbased on maximizing the distance between the hyperplane and the nearesttraining data points (support vectors). In our work, we used a radial basisfunction (RBF) as the kernel. The RBF kernel was used due to its ability fornonlinear mapping of the data. RBF outperforms a linear kernel in terms of443.3. Classification and cross-validationclassification accuracy in our data. The classifier was a soft margin classifierwhere the cost or error penalty, c, and the RBF kernel parameter, γ, werethe two parameters to tune.If the training set consists of two classes, the support vector machinesolves the following optimization problem [35]:arg minΨ,b,ζ12ΨTΨ + cN∑i=1ζi (3.9)subject to : yi(ΨTϕ(xi) + b) ≥ (1− ζi)and ζi ≥ 0where the training dataset, {x1, ...,xN} consists of N observation with itsassociated class labels {y1, ..., yN} where yi{−1,+1}. b and Ψ define the hy-perplane, ζi is the slack variable that allow for misclassification of noisy datapoints and c > 0 controls the trade-off between the error margin and slackvariable penalty. The training vectors xi are mapped to a high-dimensionalspace by the mapping function ϕ. The vector Ψ that determines the opti-mal hyperplane can be written as a linear combination of training vectors:Ψ =∑yiαiϕ(xi) [12]. So the optimization equations defined above nowcontain the mapping function in its dot product format, and this dot prod-uct can be defined by the kernel function K(xi,xj) = ϕ(xi)Tϕ(xj). In ourwork, we used a radial basis function (RBF) as the kernel defined as :K(xi,xj) = e−γ||xi−xj||2; γ > 0 (3.10)The solution to this optimization problem is through Lagrangian multipliersand transformation to a dual convex form that can be solved using quadraticprogramming.For the implementation of the classifier we used LIBSVM, which is a C++implementation of the SVM algorithm [10]. Since SVM is a non-probabilistic453.3. Classification and cross-validationbinary classifier, it cannot give class likelihood values. Therefore, in orderto evaluate the class likelihood based on SVM classification, we used themethodology proposed by Platt [71, 91]. This method estimates the classlikelihood in form of a sigmoid likelihood function of the estimated SVMhyperplane equation. The parameters of this sigmoid function were estimatedusing maximum likelihood estimation on the training data. By applying acutoff threshold to the cancer likelihood value acquired for each ROI wewere able to calculate confusion matrices and plot the Receiver OperatingCharacteristic (ROC) curves.We investigated the performances of different feature combinations inclassifying tumor tissues. Six different SVM classifiers were trained on the fol-lowing combinations of the features: the six dimensional empirical model-freeparameters, the five dimensional PCA features (consisted of the PCA featuresthat can explain 97% variance of the data), the combination of the empiri-cal model-free features with pharmacokinetic parameters, the combination ofthe five-dimensional PCA and pharmacokinetic parameters, 22-dimensionalLASSO-isolated PCA features and the three dimensional pharmacokineticfeature vector. For each of these cases, the SVM classifier was tuned bycross-validation. The cross-validation was performed on a leave-one-patient-out basis, where for each combination of c and γ the classifier was trainedon 15 patients and tested on the other patient. This process was carried outfor all of the patients with the cross-validation targeted to maximizing thetotal area under ROC. We investigated the possible combinations of c and γby a grid search on c ∈ {2−10, 2−9.5, ..., 210} and γ ∈ {2−10, 2−9.5, ..., 210}.The evaluation of the classifiers’ performances was based on the Area Un-der receiver operating characteristics Curve (AUC), sensitivity and specificityvalues. We also described a measure to quantify the classifier’s performancein detecting tumor regions, namely slice-level sensitivity. To define slice-levelsensitivity we took the prediction of the classifier as a ‘true positive’ if it cancorrectly classify more than 50% of the total tumor area in a slide, and as a463.4. ResultsTable 3.1: Statistics of the three pharmacokinetic parameters in the formatof mean (standard deviation)Ktrans ve vpNormal 0.053 (0.03) 0.187 (0.08) 0.022 (0.02)Cancer 0.104 (0.06) 0.214 (0.07) 0.018 (0.01)‘false negative’ otherwise. Hence slice-level sensitivity is defined as:Slice− level sensitivity =TPslidesTPslides + FNslides(3.11)where, TPslides = total number of slides where the classifier detected at least50% of the total tumor area,and, FNslides = total number of slides where the classifier detected less than50% of the total tumor area.3.4 ResultsWe used 449 regions of interest from 16 patients to train and test the clas-sifier. This covered the peripheral zone of the prostate gland in these cases.Out of these samples, 219 were from the cancer regions and the remaining230 were from the normal areas. These samples were extracted from theslices where the tumor was larger than 0.5 cm2 in area. The mean and thestandard deviation of the pharmacokinetic parameters, the empirical model-free parameters and the first five PCA features are shown in Tables 3.1, 3.2and 3.3.We trained six different classifiers on different sets of features. Table 3.4summarizes the classifiers’ performances. The feature sets were:• Pharmacokinetic features: Ktrans, ve, and vp• Model-free empirical features described in Section 3.2.1473.4. ResultsTable 3.2: Statistics of the six empirical model-free features in the formatof mean (standard deviation)Imax tpeak tonset ∆initial ∆mean ∆washoutNormal 1.536 5.195 0.491 0.341 0.644 -0.053(0.50) (2.75) (0.11) (0.14) (0.48) (0.09)Cancer 1.937 3.535 0.452 0.465 1.718 -0.069(0.55) (2.22) (0.08) (0.18) (1.6) (0.04)• First five PCA features• Ktrans and empirical features• Ktrans and first five PCA features• LASSO-isolated PCA features described in Section 3.2.3Pharmacokinetic parameters : When only the pharmacokinetic pa-rameters were used, the AUC was 0.78. The sensitivity and specificity was76.1% and 63.9% respectively. The pharmacokinetic parameters are ex-tracted as described in Section 2.3.2 and averaged over the region of interest.Empirical model-free parameters : With the empirical model-freeparameters, the overall Area Under the ROC (AUC) was 0.78, with a stan-dard deviation of 0.16. This was obtained when the SVM parameters werec = 0.04 and γ = 0.02. We found that in 23 out of 33 slides with peripheralzone cancer, the classifier can detect more than 50% of the tumor area, re-sulting in a slice-level sensitivity of 70%, when the classifier was trained onthe empirical parameters.First five PCA parameters : When the classifier was trained on thefirst five PCA features alone, with c = 0.7 and γ = 0.001, the AUC wasfound to be 0.79 with a standard deviation of 0.18. For a cutoff value of 0.43applied to the cancer likelihood, the sensitivity and specificity were 71.2%and 72.2% respectively. The slice-level sensitivity was 70%.483.4. ResultsTable 3.3: Statistics of the first five PCA features in the format of mean(standard deviation)feature-1 feature-2 feature-3 feature-4 feature-5Normal 5.122 (1.76) 0.461 (0.52) -0.269 (0.09) -0.227 (0.12) -0.078 (0.08)Cancer 6.204 (1.7) 1.245 (0.81) -0.314 (0.11) -0.259 (0.18) -0.043 (0.14)Ktrans and empirical parameters : We also examined the cancer de-tection capability of the DCE method when the pharmacokinetic parameterswere used along with the model-free parameters. We trained a classifier withKtrans and model-free empirical parameters together. Adding ve and vp tothis set did not improve the result and therefore we included only Ktrans inthis analysis. The optimum parameters found for the SVM classifier were c= 0.06 and γ = 0.008 and the area under ROC curve was 0.8 with a standarddeviation of 0.14. The slice-level sensitivity was found to be 73%, meaningthat in 73% of the slices with a pathologic finding, more than 50% of thetumor area was classified as cancer.Ktrans and first five PCA parameters : When the classifier wastrained on the first five PCA features along with Ktrans, with c = 0.5 andγ = 0.001, the AUC was 0.8 with a standard deviation of 0.14. Note thatthe increase, compared to only using PCA, was not statistically significant.At the cutoff value of 0.42 for the class likelihood, 163 tumor samples werecorrectly classified while 56 tumor samples were missed. The sensitivity andspecificity were found to be 74.4% and 70.9% respectively. The slice-levelsensitivity was 75.8%.LASSO-isolated PCA parameters : We ranked the PCA featuresbased on the magnitude of their corresponding coefficient in LASSO regres-sion and selected the optimal number of features by forward search and leave-one-patient-out cross-validation. The maximum AUC was obtained when 22PCA features isolated by LASSO were used to train the classifier. With493.4. ResultsTable 3.4: Area under receiver operating characteristic curve (AUC), sensitiv-ity, specificity and slice-level sensitivity with different feature combinations.Slice-level sensitivity is defined as the percentage of cases where the classifiercan detect more than 50% of the total tumor area. AUC is enlisted in theformat of mean (standard deviation).Features AUC Sensitivity SpecificitySlice-levelsensitivityPharmacokinetic Parameters 0.78 (0.14) 76.1% 63.9% 78.8%Model-free Empirical Parameters 0.78 (0.16) 75.3% 64.0% 70.0%Ktrans and Empirical Parameters 0.80 (0.14) 75.3% 67.8% 73.0%PCA features 0.79 (0.18) 71.2% 72.2% 70.0%Ktrans and PCA Features 0.80 (0.14) 74.4% 70.9% 75.8%LASSO-isolated PCA Features 0.86 (0.11) 81.5% 77.5% 90.9%the LASSO-isolated PCA components of the T1-weighted intensities, we ob-tained an AUC of 0.86, with c = 1 and γ = 0.008. At the cutoff valueof 0.42, the sensitivity and specificity were 81.5% and 77.5% respectively.The slice-level sensitivity of 90.9%. This combination resulted in the high-est area under ROC and slice-level sensitivity among the different featurecombinations that were studied. Even though the improvement was not sta-tistically significant, LASSO-isolated features are computationally easier andtheir performance was as good as pharmacokinetic features.We used the classifier trained on LASSO-isolated PCA features to gen-erate cancer likelihood maps for the entire peripheral zone of the gland. Togenerate cancer likelihood maps, the classifier was trained on all other pa-tients. Then each pixel from the DCE MRI image of interest was used as atest sample for the classifier. Then LASSO-isolated PCA components wereextracted for each pixel and the predicted cancer likelihood of the classifierwas mapped onto the T1-weighted image as a cancer likelihood map. Fig-ures 3.4 to 3.7 show the cancer likelihood maps for five cases generated withthis method. The likelihood maps were plotted using standard Jet colormap,503.4. Results3+4(a) (b) (c)Figure 3.4: (a) Peripheral zone tumor marked in pathology slide. (b) Pathol-ogy slide registered to corresponding DCE image. (c) The generated cancerlikelihood map superimposed on the DCE T1-weighted image. Note that theclassifier is only trained on the peripheral zone. The classifier is trained onall other cases.3+4(a) (b) (c)Figure 3.5: (a) Peripheral zone tumor marked in pathology slide. (b) Pathol-ogy slide registered to corresponding DCE image. (c) The generated cancerlikelihood map superimposed on the DCE T1-weighted image. Note that theclassifier is only trained on the peripheral zone. The classifier is trained onall other cases.513.4. Results3+4(a) (b) (c)Figure 3.6: (a) Peripheral zone tumor marked in pathology slide. (b) Pathol-ogy slide registered to corresponding DCE image. (c) The generated cancerlikelihood map superimposed on the DCE T1-weighted image. Note that theclassifier is only trained on the peripheral zone. The classifier is trained onall other cases.(a) (b) (c)Figure 3.7: (a) Peripheral zone tumor marked in pathology slide. (b) Pathol-ogy slide registered to corresponding DCE image. (c) The generated cancerlikelihood map superimposed on the DCE T1-weighted image. Note that theclassifier is only trained on the peripheral zone. The classifier is trained onall other cases.523.5. Summarywhere hot colors represent higher cancer likelihood values. As can be seen,the cancer likelihood maps show higher cancer likelihood of cancer in thetumor area.3.5 SummaryIn this chapter, we have developed a novel data-driven and model-free frame-work for prostate cancer detection using DCE MRI. We report an area un-der ROC curve of 0.86 with our proposed LASSO-isolated PCA featuresextracted from the DCE T1-weighted intensities. The comparisons in thischapter show that the LASSO-isolated features outperform both the quan-titative pharmacokinetic modeling, and the currently available model-freemethods such as empirical analysis.The proposed data-driven approach to prostate cancer detection fromDCE-MR images shows an improvement compared to the traditional phar-macokinetic modeling. Even though the improvement is not statisticallysignificant, most likely due to our limited sample size, the data-driven ap-proach removes the need for AIF calculation and modeling. Our method alsogenerates helpful cancer likelihood maps for MRI-targeted biopsies. Boththe empirical model-free parameterization of the normalized T1-weighted in-tensity and the principal component analysis of the T1-weighted intensitiesresult in features for cancer detection that are effective, in terms of AUC,when compared with the standard pharmacokinetic parameters. This novelapproach can significantly simplify and streamline the use of DCE imagingfor prostate cancer detection.53Chapter 4Multiparametric MRI featureextraction and classification4.1 IntroductionMultiparametric Magnetic Resonance Imaging (mpMRI), where parametersfrom different MRI modalities are used in combination is now considered partof the standard of care for image-based evaluation of prostate to determinethe need for a biopsy in many parts of the world [5]. Numerous studiesreported improved accuracy of prostate cancer detection and localizationwhen different MR modalities are used together. Among other modalities,Diffusion Weighted Imaging (DWI) has the ability to characterize the de-phasing of MR signal caused by molecular diffusion. In cancerous regions,the regular tissue distribution pattern in prostate is disturbed and replacedby masses of malignant epithelial cells and glands. This pathological changealters tissue diffusion pattern in prostate that results in changing parametervalues extracted from diffusion MRI.Diffusion Tensor Imaging (DTI) is an advanced form of DWI that en-ables the measurement of directionality along with the magnitude of waterdiffusion. While DWI generates one diffusion parameter, namely ApparentDiffusion Coefficient (ADC), most of the studies on quantitative DTI usestwo diffusion parameters, namely average diffusivity (〈D〉) and FractionalAnisotropy (FA). ADC, which is extracted from DWI, denotes the averagediffusivity in tissues. On the other hand, diffusivity extracted from DTI isa tensor and the trace of the diffusion tensor is denoted by 〈D〉. FA is an544.1. Introductionindication of how anisotropic the diffusion process is and it can only be ex-tracted from DTI. Decreased ADC and 〈D〉 values are reported frequentlyas a strong indicator of tumors [26, 37, 93]. Gibbs et al. [27] and Wang etal. [90] reported inverse relationships between diffusivity values and tumorproliferation using histological measurements of cellular density. Hambrocket al. [29] and Tamada et al. [84] reported negative correlation betweendiffusivity values and Gleason grades. However, the association of FA withtumor is indecisive and different studies reported increased [26], decreased[55] and even unchanged values of FA in prostate carcinoma [92].Several studies reported that when quantitative parameters extractedfrom diffusion MRI and DCE MRI are used together, it results in a bet-ter cancer detection accuracy than when used separately [13, 31]. Oto et al.[64] reported significant increase in sensitivity when Ktrans is used with aver-age diffusivity values with multivariate logistic regression. Delongchamps etal. [14] reported significantly better cancer detection performance in periph-eral zone of prostate when DCE and diffusion MRI are used together withT2 compared to when DCE and diffusion MRI are used separately with T2.Langer et al. investigated different combinations of parameters from DCE,DTI and T2 and the highest performance in terms of area under ReceiverOperating Characteristics curve (ROC) was achieved when 〈D〉 and Ktranswere used with T2 [46]. In [47], these three parameters were reported tobe significantly correlated with specific histologic components that differ be-tween normal and cancerous peripheral zone of prostate, and hence can beused in combination as an image-based prognostic parameter.Some studies focused on applying machine learning approaches to sepa-rate cancer and normal tissues by classifying multiparametric MRI param-eters [9, 65]. Most of these studies reported their performances in terms ofarea under ROC, sensitivity, specificity and accuracy. In Kozlowski et al.[43], it is reported that the combination of DTI and DCE parameters at 3Tesla results in improved cancer diagnostic capability in terms of area under554.2. Feature extraction and classificationROC. Moradi et al. also reported improved performance with DCE and DTIfeatures and generated a single parameter map of cancer likelihood usingsupport vector machine classification [61].In this chapter, we investigated the performance of data-driven DCE fea-tures in a multiparametric framework. In Chapter 3, we proposed a set ofdata-driven features generated by principal component analysis of DCE T1time course and reduced its dimensionality by a Least Absolute Shrinkageand Selection Operator (LASSO). In this chapter, this set of DCE featuresis used with DTI features extracted from the registered diffusion images fortumour detection using Support Vector Machine (SVM) classification. Themethod was validated in 16 clinical cases based on wholemount histopathol-ogy slides as the reference. Using this computational framework, we showthat the proposed approach to parametrization of DCE data can improve thedetection of cancer from DCE data within the mpMRI protocol.4.2 Feature extraction and classificationIn this work, we extracted model-free PCA features from DCE-MRI and dif-fusivity, 〈D〉 from DTI. The features are extracted from the multiparametricMRI data of 16 patients scheduled for radical prostatectomy described inSection 2.2. The normal and cancer regions were mapped from the whole-mount pathology slides to the corresponding DCE and DTI slides by theregistration algorithm discussed in Section 2.4 and the mapped regions werethen used to extract features. To extract features, each tumor was takenas a region of interest (ROI) and tumors larger than 100 mm2 were dividedinto two or more smaller ROIs. In the training set, we only considered thosetumors that were larger than 47 mm2. We considered the whole tumor as anROI to minimize the effect of registration inaccuracies in ROI selection. Notethat this is different from Chapter 3 where we only used DCE image-basedfeatures, and hence no registration was necessary to extract features.564.3. ResultsFrom the dynamic T1-weighted MR images, the intensity values withineach ROI were averaged and an average time course signal was formed. Ourapproach to data-driven characterization of the DCE time course is to usea dimensionality reduction method to convert the time series of normalizedcontrast enhanced T1-weighted intensities to an optimally sized vector offeatures. We used the method described in Sections 3.2.2 and 3.2.3 for thispurpose, and extracted the most significant PCA components. The value ofλ for LASSO was determined by 10-fold cross validation on the entire datasettargeted to minimize the mean squared error.From diffusion tensor images, we extracted two parameters: average dif-fusivity (〈D〉) and Fractional Anisotropy (FA). 〈D〉 is a measure of averagediffusivity and FA measures how anisotropic the diffusion process is. Thefeatures extracted from DCE and DTI were then used together as a multi-parametric MRI (mpMRI) based feature vector and the true label (class) ofthe feature vector was determined from the pathology image.Soft margin Support Vector Machine (SVM) classifier (described in Sec-tion 3.3) was used to classify cancer and normal tissues with mpMRI featurecombination. The margin violation penalty weight, c, and the Radial BasisFunction (RBF) kernel parameter, γ, were the two parameters to tune. Wetrained two different classifiers- one with traditional multiparametric features(three pharmacokinetic and two DTI features) and one with the combined〈D〉 and LASSO-isolated PCA features. We did not include FA in our pro-posed mpMRI feature set as it did not improved the result significantly. Foreach of these cases, the classifier was tuned by cross-validation on a leave-one-patient-out basis. We investigated the possible combinations of c and γby a grid search on c ∈ {2−10, 2−9.5, ..., 210} and γ ∈ {2−10, 2−9.5, ..., 210}, andthe cross-validation was targeted at maximizing the AUC.574.3. ResultsTable 4.1: Statistics of the traditinal multiparametric MRI features in theformat of mean (standard deviation).〈D〉FAKtransve vp(10−3 mm2/s) (min−1)Peripheral ZoneNormal1.616 0.165 0.054 0.190 0.021(0.36) (0.05) (0.02) (0.07) (0.01)Cancer1.468 0.175 0.102 0.218 0.016(0.22) (0.04) (0.04) (0.05) (0.01)Central GlandNormal1.485 0.152 0.090 0.245 0.023(0.23) (0.04) (0.04) (0.07) (0.02)Cancer1.436 0.176 0.115 0.266 0.020(0.16) (0.04) (0.05) (0.07) (0.02)4.3 ResultsAt first, we investigated the performance of the data-driven mpMRI featuresin the peripheral zone only. We used 191 regions of interest from the periph-eral zone of 16 patients to train and test the classifier. Out of these samples,92 were from cancer regions and the remaining 99 were from normal areas.These samples were extracted from the slices where the tumor was largerthan 0.5 cm2 in area.At first, we trained a classifier with traditional multiparamteric features,i.e. 〈D〉 and FA from diffusion MRI and Ktrans, ve and vp from DCE MRI.Table 4.1 shows the statistics of the traditional mpMRI features used in thisstudy. With the traditional parameters, the overall AUC was 0.80, with astandard deviation of 0.12. The optimal threshold was 0.24, and at thisthreshold, the sensitivity and specificity was 73.9% and 73.7% respectively.584.3. ResultsTable 4.2: Area under receiver operating characteristic curve (AUC), sensitiv-ity, specificity and slice-level sensitivity with different feature combinations.Slice-level sensitivity is defined as the percentage of cases where the classifiercan detect more than 50% of the total tumor area.Features AUC Sensitivity SpecificitySlice-levelsensitivityPeripheral Zone ClassifierTraditional 0.8073.9% 73.7% 74.3%mpMRI Features (0.12)〈D〉 and 0.9185.9% 80.8% 91.4%LASSO-PCA Features (0.14)Whole Gland ClassifierTraditional 0.6862.2% 63.5% 60.0%mpMRI Features (0.2)〈D〉 and 0.8780.7% 82.0% 85.0%LASSO-PCA Features (0.15)The slice-level sensitivity was defined as percent of cases where the classifiercan detect more than 50% of the total tumor area. With the traditionalfeatures, the slice-level sensitivity was 74.3%, where in 26 out 35 cases theclassifier detected more than 50% of the tumor area.When the classifier was trained on the LASSO-isolated PCA featuresalong with 〈D〉 from diffusion MRI, the area under ROC was 0.91, witha standard deviation of 0.14. The sensitivity, specificity and accuracy was85.9%, 80.8% and 83.2% respectively. In 32 out of 35 slices, the classifierdetected more that 50% area of the tumor, resulting in a slice-level sensitivityof 91.4%. The statistical significance test between the AUC values with theproposed features and traditional mpMRI features generated a p-value of 0.5.We also investigated the performance of the proposed data-driven mpMRI594.3. Resultsfeature combination in detecting cancer from the entire prostate gland. Weextracted 111 regions of interest (43 tumor, 68 normal) from the centralgland, and combined these samples with the peripheral zone samples. In totalthere were 302 regions of interest, with 135 tumor samples and 167 normalsamples from the whole gland, and these samples were used to train and testthe whole gland classifiers. With traditional multiparametric features, theAUC over 16 patients was 0.68. At the threshold of 0.42, the sensitivity andspecificity were 62.2% and 63.5% respectively. In 24 out of 40 slices, theclassifier detected more than 50% of the total tumor area, and the slice-levelsensitivity was 60.0%. The classifier was trained on leave-one-patient-outbasis.When the proposed LASSO-isolated PCA features were used with 〈D〉,the area under ROC curve was 0.87, with a standard deviation of 0.15. Thenumber of LASSO-isolated PCA features to be used in the analysis was de-termined by forward search algorithm targeted to maximize the AUC. Atthe optimal threshold of 0.5, the sensitivity, specificity and accuracy were80.7%, 82.0% and 81.5% respectively. In 34 out of 40 slices, the classifier de-tected more than 50% area of the tumor, and hence the slice-level sensitivitywas 85.0%. The statistical significance test between the AUC values withthe proposed features and traditional mpMRI features generated a p-valueof 0.4. Table 4.2 summarizes the peripheral zone and whole gland classifiers’performances.We analysed the cancer likelihood values generated by the peripheral-zoneclassifier trained on our proposed multiparametric features to find associationwith the stage of cancer. Table. 4.3 shows the mean cancer likelihood valueswith their corresponding Gleason scores. The mean value of cancer likelihoodfor tumors with Gleason score (3+3) and (3+4) was 0.651, calculated over77 samples. For more aggressive tumors the average cancer likelihood valuewas higher than that calculated for less aggressive tumors. We had 15 sam-ples for aggressive tumors with Gleason score (3+4+5) and (4+3), and the604.3. ResultsTable 4.3: Generated average cancer likelihood values in the format of mean(standard deviation) with different Gleason scores. The likelihood scoreswere calculated using the peripheral-zone classifier trained on the proposedmpMRI features.Normal (3+3) and (3+4+5) andsamples (3+4) tumors (4+3) tumorsNumber of samples 99 77 15Average cancer likelihood0.146 0.651 0.780(0.19) (0.34) (0.27)mean likelihood value for these samples was 0.78. For normal samples themean likelihood value was 0.146. To calculate the correlation coefficient weassigned a score based on the Gleason score to the samples. The correlationcoefficient between the predicted likelihood and the Gleason sum is found tobe 0.69. Since we have few samples from (3+4+5) and (4+3) tumors, wegrouped them together and assigned a Gleason sum of (3+5) to calculate thecorrelation coefficient.We used the classifier trained on the proposed mpMRI features (〈D〉 andLASSO-isolated PCA features) to generate cancer likelihood maps for theprostate gland. To generate cancer likelihood maps, the classifier was trainedon all other patients and each pixel from the image of interest was used asa test sample for the classifier. To extract the features, each pixel from theT2-weighted MR image was mapped to corresponding DTI and DCE MRimage, and features were extracted from the mapped pixels. The predictedcancer likelihood of the classifier was mapped onto the T2-weighted MRI asa single parameter map of cancer likelihood.Figures 4.1 to 4.3 are the cancer likelihood maps generated for the pe-ripheral zone only. The peripheral zone classifier trained on 〈D〉 and LASSO-isolated PCA features was used to generate these cancer likelihood maps. In614.3. Results3+4(a) (b) (c)Figure 4.1: (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the corresponding T2-weightedimage. (c) The generated cancer likelihood map superimposed on the T2-weighted image. Note that the classifier is only trained on the peripheralzone. The classifier is trained on all other cases.(a) (b) (c)Figure 4.2: (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the corresponding T2-weightedimage. (c) The generated cancer likelihood map superimposed on the T2-weighted image. Note that the classifier is only trained on the peripheralzone. The classifier is trained on all other cases.624.3. Results(a) (b) (c)Figure 4.3: (a) Peripheral zone tumor marked in pathology slide for onepatient. (b) Pathology slide registered to the corresponding T2-weightedimage. (c) The generated cancer likelihood map superimposed on the T2-weighted image. Note that the classifier is only trained on the peripheralzone. The classifier is trained on all other cases.each of the cases, the findings of the likelihood maps are consistent with thepathology images.We also generated cancer likelihood maps for the entire prostate usingthe whole gland classifier trained on 〈D〉 and LASSO-isolated features. Fig-ures 4.4 and 4.5 show the cancer likelihood maps generated for the entireprostate gland. We generated two cancer likelihood maps: one from thempMRI features and one from the DCE features alone, to observe the effectof registration inaccuracy. As can be seen, the generated cancer likelihoodmaps detected high likelihood of tumor in the same areas where tumor wasoutlined in the pathology images. However, due to inaccuracy of the reg-istration algorithm, the tumor regions are deformed in the likelihood mapsgenerated from mpMRI features. As discussed earlier in Chapter 2, in somecases different degrees of deformation is observed in the central gland andin the peripheral zone, whereas B-spline transformation deforms the wholeprostate gland simultaneously. This limitation of the registration algorithmresulted in deformation of the tumor regions in the cancer likelihood maps.Furthermore, while generating the cancer likelihood maps, we mapped each634.3. Results(a)(b) (c)(d) (e)Figure 4.4: (a) Peripheral zone tumor marked in pathology slide for one pa-tient. (b) Corresponding DCE image. (c) Cancer likelihood map generatedusing the classifier trained on data-driven DCE features only. (d) Corre-sponding T2-weighted image. (e) Cancer likelihood map generated using theclassifier trained on the proposed mpMRI features, registered to T2-weightedimage.644.3. Results(a)(b) (c)(d) (e)Figure 4.5: (a) Peripheral zone tumor marked in pathology slide for one pa-tient. (b) Corresponding DCE image. (c) Cancer likelihood map generatedusing the classifier trained on data-driven DCE features only. (d) Corre-sponding T2-weighted image. (e) Cancer likelihood map generated using theclassifier trained on the proposed mpMRI features, registered to T2-weightedimage.654.4. Summarypixel to its corresponding DTI and DCE images, and features were extractedfrom the mapped pixels. However, the registration algorithm proposed heredoes not have pixel-level accuracy. Hence discrepancy is observed betweentumor areas in cancer likelihood maps generated from mpMRI features andDCE features alone, where no registration was necessary to generate likeli-hood maps.Another limitation of the generated whole gland cancer likelihood mapsis that, regions of false positives are observed in the central gland. In mostof the cases, the regions near urethra are erroneously detected as tumor, inboth the cancer likelihood maps generated by mpMRI features and DCEfeatures alone, as can be seen in Figure 4.5. This might be due to the factthat the tissue pattern is different in these regions, and one single classifierfor the whole gland might not be the best tool to generate cancer likelihoodmaps. It might be helpful to train zone-specific classifiers and combine theiroutcomes to generate whole gland cancer likelihood maps.4.4 SummaryIn this chapter, we showed that the data-driven parametrization of the nor-malized T1-weighted signal intensities using PCA and LASSO, results infeatures for cancer detection that are effective, in terms of AUC, in a mul-tiparametric framework. They provide an improved AUC when used in anmpMRI protocol that included diffusion MRI feature. We reported an AUCof 0.87 when the LASSO-isolated PCA parameters were employed with DTIfeature for the whole gland classification, and an AUC of 0.91 for the pe-ripheral zone-limited classification. We showed correlation of the generatedcancer likelihood scores with their corresponding Gleason scores. Cancer like-lihood maps were also generated for the whole prostate gland that showedhigher cancer likelihood in cancerous regions. This shows the potential ofcombining the diffusion MR parameter with the data-driven DCE parame-664.4. Summaryters as an imaging biomarker in prostate cancer diagnosis.67Chapter 5Conclusions5.1 SummaryIn this thesis, we developed a complete image processing pipeline of prostatecancer detection from dynamic contrast enhanced and diffusion tensor im-ages, starting from the registration of DTI and DCE MR images to theircorresponding pathology images, followed by extraction of data-driven fea-tures, and finally generation of a single parameter map of cancer likelihood.The major contributions of this work are:• A semi-automatic pharmacokinetic parameter extraction pipeline fromdynamic contrast enhanced MR images is proposed. The framework issemi-automatic and includes vessel extraction by circular Hough trans-form to calculate arterial input function and uses extended Kety modelto generate the pharmacokinetic parameter maps.• An image-based technique of prostate cancer detection is proposedthat is based on the novel image-based features extracted from DCET1-weighted intensities. It is shown that the proposed framework oflearning the tissue signature directly from T1-weighted intensities isa promising approach for cancer detection from DCE MRI data aloneand in multiparametric settings.• Based on the proposed data-driven technique, a single parameter mapof cancer likelihood is generated that has the potential to be used as areference for MRI-targeted biopsies.685.2. Discussions and limitations5.2 Discussions and limitationsThe major limitation of the dataset used in this work is the small samplesize and low temporal resolution of the DCE MRI. Our dataset consists of 16patients and it is difficult to interpret the results and find association withquantitative parameters with such a small sample set.Another limitation is that the proposed approach is not independent ofMRI technique used. Since both the model-free empirical and the PCA anal-ysis rely on the normalized image intensity, the results could depend on thepulse sequence parameters. On the other hand, pharmacokinetic parametersare calculated from the contrast agent concentration values and are quan-titative physical parameters. Therefore, they should be, theoretically, lessdependent on the pulse sequence parameters. As discussed in the introduc-tion, this theoretical advantage of the pharmacokinetic parameters is nottranslated into a clinical advantage due to the sensitivity of the calculationprocess to noise and to the accurate extraction of the AIF. Even thoughthe pharmacokinetic modeling provides results that can be pulse sequenceindependent, clinically the most important issue is the ability to detect can-cer, which appears to be improved with the use of the model-free approachdemonstrated here.The registration methodology employed here has several shortcomings.It deforms the whole prostate simultaneously whereas in some cases differentdegrees of deformation is needed for the peripheral zone and the centralgland. This limitation resulted in deformation of tumor in the likelihoodmaps. Also, the registration methodology does not have pixel level accuracy,hence the pixel level cancer likelihood maps showed false positives in somecases.It should be noted that the proposed data-driven parametrization methoddoes not replace accurate physical modeling of the perfusion process. Thosemodels provide physical and physiological insights for understanding the pro-cess of cancer growth. Here, we show that our data-driven method provides695.3. Future workimproved diagnosis. Although the differences between enhancement curvesfrom normal and cancerous tissue analysed by the PCA come from differ-ences in perfusion patterns, the model-free empirical or the PCA method initself does not provide any insight into these differences.In this work, we did not use pixel resolution for feature extraction intraining set to ensure that the training set does not get trained on data thatis affected by registration error or imaging noise. Averaging over a region ofinterest makes the classifier less sensitive to noisy pixels. Also, our target isto detect clinically significant prostate cancer. A prostate tumor is acceptedas clinically significant if its volume exceeds 0.5 mL [31]. However, whilegenerating cancer likelihood maps, we used each pixel as a test sample forthe trained classifier and predicted the cancer likelihood values for each pixel.5.3 Future workThis thesis work opens up some exciting fields to focus on for future improve-ments. Some of the areas can be as follows:• One future development on this approach can be to investigate theassociation of the proposed features with actual physical parameters,i.e. the pharmacokinetic parameters, tissue microvasculature, etc.• Another further improvement can be to overcome the limitations ofthe registration framework. As discussed in Chapter 4 Section 4.3, theinaccuracy of the registration framework resulted in tumor deformationand false positives in the cancer likelihood maps. One solution mightbe to register the peripheral zone and central gland separately.• 3-D registration framework can be implemented for pathology slidesto MR registration and for T2-weighted to DCE/DTI registration asopposed to 2-D framework described in this work.705.3. Future work• 7T MR images can be acquired and used as an intermediate modal-ity for registering high resolution pathology to low resolution 3T MRimages.• Features from other MR modalities can be used in combination withthe proposed data-driven features. For example- adding quatitative T2features or MR elastography features can improve the cancer detectionability of the proposed classifiers.71Bibliography[1] H. Abdi and L. J. Williams. Principal Component Analysis. WileyInterdisciplinary Reviews: Computational Statistics, 2(4):433–459, 2010.[2] H. Augustin, G. Fritz, T. Ehammer, M. Auprich, and K. Pummer. Accu-racy of 3-Tesla Magnetic Resonance Imaging for the staging of prostatecancer in comparison to the Partin tables. Acta Radiologica, 50(5):562–569, 2009.[3] D. H. Ballard. Generalizing the Hough transform to detect arbitraryshapes. Pattern recognition, 13(2):111–122, 1981.[4] J. O. Barentsz, M. Engelbrecht, G. J. Jager, J. A. Witjes,J. de LaRosette, van der Sanden, et al. Fast dynamic gadolinium-enhanced MR imaging of urinary bladder and prostate cancer. Journalof Magnetic Resonance Imaging, 10(3):295–304, 1999.[5] J. O. Barentsz, J. Richenberg, R. Clements, P. Choyke, S. Verma,G. Villeirs, et al. ESUR prostate MR guidelines 2012. European So-ciety of Urogenital Radiology, 22(4):746–757, 2012.[6] B. N. Bloch, N. M. Rofsky, R. H. Baroni, R. P. Marquis, I. Pedrosa, andR. E. Lenkinski. 3 Tesla Magnetic Resonance Imaging of the prostatewith combined pelvic phased-array and endorectal coils: Initial experi-ence1. Academic Radiology, 11(8):863–867, 2004.[7] D. L. Buckley, C. Roberts, G. J. Parker, J. P. Logue, and C. E. Hutchin-son. Prostate Cancer: Evaluation of Vascular Characteristics with Dy-72Bibliographynamic Contrast-Enhanced T1-weighted MR Imaging - Initial Experi-ence. Radiology, 233(3):709–715, 2004.[8] Canadian Cancer Society. Prostate cancer statistics. http://www.cancer.ca/en/cancer-information/cancer-type/prostate/prostate-cancer/?region=bc.[9] I. Chan, W. Wells III, R. V. Mulkern, S. Haker, J. Zhang, K. H. Zou,et al. Detection of prostate cancer by integration of line-scan diffusion,T2-mapping and T2-weighted magnetic resonance imaging; a multichan-nel statistical classifier. Medical Physics, 30:2390, 2003.[10] C. C. Chang and C. J. Lin. LIBSVM: A library for support vectormachines. ACM Transactions on Intelligent Systems and Technology, 2(3):27:1–27:27, 2011.[11] J. Chen, J. Yao, and D. Thomasson. Automatic determination ofarterial input function for dynamic contrast enhanced MRI in tu-mor assessment. In Medical Image Computing and Computer-AssistedIntervention–MICCAI 2008, pages 594–601. Springer, 2008.[12] C. Cortes and V. Vapnik. Support-vector networks. Machine Learning,20(3):273–297, 1995.[13] M. de Rooij, E. H. J. Hamoen, J. J. Fu¨tterer, J. O. Barentsz, andM. M. Rovers. Accuracy of Multiparametric MRI for Prostate CancerDetection: A Meta-Analysis. American Journal of Roentgenology, 202(2):343–351, 2014.[14] N. B. Delongchamps, F. Beuvon, D. Eiss, T. Flam, N. Muradyan,M. Zerbib, et al. Multiparametric MRI is helpful to predict tumor focal-ity, stage, and size in patients diagnosed with unilateral low-risk prostatecancer. Prostate Cancer and Prostatic Diseases, 14(3):232–237, 2011.73Bibliography[15] B. Djavan, A. Zlotta, C. Kratzik, M. Remzi, C. Seitz, C. C. Schulman,and M. Marberger. PSA, PSA density, PSA density of transition zone,free/total PSA ratio, and PSA velocity for early detection of prostatecancer in men with serum PSA 2.5 to 4.0 ng/mL. Urology, 54(3):517–522, 1999.[16] B. Drew, E. C. Jones, S. Reinsberg, A. C. Yung, S. L. Goldenberg,and P. Kozlowski. Device for sectioning prostatectomy specimens tofacilitate comparison between histology and in vivo MRI. Journal ofMagnetic Resonance Imaging, 32(4):992–996, 2010.[17] W. Ellis and M. Brawer. The significance of isoechoic prostatic carci-noma. The Journal of Urology, 152(6 Pt 2):2304–2307, 1994.[18] W. J. Ellis and M. K. Brawer. Repeat prostate needle biopsy: who needsit? The Journal of urology, 153(5):1496–1498, 1995.[19] M. R. Engelbrecht, H. J. Huisman, R. J. Laheij, G. J. Jager, G. J.van Leenders, et al. Discrimination of prostate cancer from normalperipheral zone and central gland tissue by using dynamic contrast-enhanced MR imaging. Radiology, 229(1):248–254, 2003.[20] E. Eyal and H. Degani. Model-based and model-free parametric analysisof breast dynamic-contrast-enhanced MRI. NMR in Biomedicine, 22(1):40–53, 2009.[21] E. Eyal, B. Bloch, E. Furman-Haran, E. Genega, W. Dewolf, G. Bubley,et al. Pattern and model based analysis of dynamic contrast enhancedprostate MRI data. In Proc Intl Soc Mag Reson Med, volume 15, page523, 2007.[22] E. Eyal, B. N. Bloch, N. M. Rofsky, E. Furman-Haran, E. M. Genega,R. E. Lenkinski, and H. Degani. Principal component analysis of dy-74Bibliographynamic contrast enhanced MRI in human prostate cancer. InvestigativeRadiology, 45(4):174–181, 2010.[23] J. Ferlay, E. Steliarova-Foucher, J. Lortet-Tieulent, S. Rosso, J. Co-ebergh, H. Comber, et al. Cancer incidence and mortality patterns inEurope: estimates for 40 countries in 2012. European Journal of Cancer,49(6):1374–1403, 2013.[24] J. Friedman, T. Hastie, and R. Tibshirani. Regularization paths forgeneralized linear models via coordinate descent. Journal of StatisticalSoftware, 33(1):1–22, 2010.[25] J. J. Fu¨tterer, S. W. Heijmink, T. W. Scheenen, G. J. Jager, C. A.Hulsbergen-Van de Kaa, J. A. Witjes, and J. O. Barentsz. ProstateCancer: Local Staging at 3-T Endorectal MR Imaging - Early Experi-ence. Radiology, 238(1):184–191, 2006.[26] P. Gibbs, M. D. Pickles, and L. W. Turnbull. Diffusion imaging of theprostate at 3.0 Tesla. Investigative Radiology, 41(2):185–188, 2006.[27] P. Gibbs, G. P. Liney, M. D. Pickles, B. Zelhof, G. Rodrigues, and L. W.Turnbull. Correlation of ADC and T2 measurements with cell densityin prostate cancer at 3.0 Tesla. Investigative Radiology, 44(9):572–576,2009.[28] M. A. Haider, T. H. van der Kwast, J. Tanguay, A. J. Evans, A.-T.Hashmi, G. Lockwood, and J. Trachtenberg. Combined T2-weightedand diffusion-weighted MRI for localization of prostate cancer. AmericanJournal of Roentgenology, 189(2):323–328, 2007.[29] T. Hambrock, D. M. Somford, H. J. Huisman, I. M. van Oort, J. A.Witjes, H. van de Kaa, et al. Relationship between apparent diffusioncoefficients at 3.0-T MR imaging and Gleason grade in peripheral zoneprostate cancer. Radiology, 259(2):453–461, 2011.75Bibliography[30] R. J. Harris. A primer of multivariate statistics. Lawrence ErlbaumAssociates Publishers, 2001.[31] J. V. Hegde, R. V. Mulkern, L. P. Panych, F. M. Fennessy, A. Fedorov,S. E. Maier, and C. Tempany. Multiparametric MRI of prostate cancer:An update on state-of-the-art techniques and their performance in de-tecting and localizing prostate cancer. Journal of Magnetic ResonanceImaging, 37(5):1035–1054, 2013.[32] C. M. Hoeks, J. O. Barentsz, T. Hambrock, D. Yakar, D. M. Somford,S. W. Heijmink, et al. Prostate cancer: multiparametric MR imagingfor detection, localization, and staging. Radiology, 261(1):46–66, 2011.[33] H. Hricak, G. Dooms, R. B. Jeffrey, A. Avallone, D. Jacobs, W. Benton,et al. Prostatic carcinoma: staging by clinical assessment, CT, and MRimaging. Radiology, 162(2):331–336, 1987.[34] H. Hricak, P. L. Choyke, S. C. Eberhardt, S. A. Leibel, and P. T.Scardino. Imaging Prostate Cancer: A Multidisciplinary Perspective.Radiology, 243(1):28–53, 2007.[35] C. W. Hsu, C. C. Chang, and C. J. Lin. A practical guide to supportvector classification, 2003.[36] M. Inahara, H. Suzuki, H. Nakamachi, N. Kamiya, M. Shimbo,A. Komiya, et al. Clinical evaluation of transrectal power dopplerimaging in the detection of prostate cancer. International urology andnephrology, 36(2):175–180, 2004.[37] B. Issa. In vivo measurement of the apparent diffusion coefficient in nor-mal and malignant prostatic tissues using echo-planar imaging. Journalof Magnetic Resonance Imaging, 16(2):196–200, 2002.76Bibliography[38] A. Jemal, M. M. Center, C. DeSantis, and E. M. Ward. Global patternsof cancer incidence and mortality rates and trends. Cancer EpidemiologyBiomarkers & Prevention, 19(8):1893–1907, 2010.[39] A. Jemal, F. Bray, M. M. Center, J. Ferlay, E. Ward, and D. Forman.Global cancer statistics. CA: A Cancer Journal for Clinicians, 61(2):69–90, 2011.[40] J. K. Kim, S. S. Hong, Y. J. Choi, S. H. Park, H. Ahn, C. S. Kim,and K. S. Cho. Wash-in rate on the basis of dynamic contrast-enhancedMRI: Usefulness for prostate cancer detection and localization. Journalof Magnetic Resonance Imaging, 22(5):639–646, 2005.[41] S. Klein, J. P. Pluim, M. Staring, and M. A. Viergever. Adaptivestochastic gradient descent optimisation for image registration. Inter-national Journal of Computer Vision, 81(3):227–239, 2009.[42] S. Klein, M. Staring, K. Murphy, M. A. Viergever, and J. P. Pluim.Elastix: a toolbox for intensity-based medical image registration. Med-ical Imaging, IEEE Transactions on, 29(1):196–205, 2010.[43] P. Kozlowski, S. D. Chang, R. Meng, B. Ma¨dler, R. Bell, E. C. Jones,and S. L. Goldenberg. Combined prostate diffusion tensor imaging anddynamic contrast enhanced MRI at 3T quantitative correlation withbiopsy. Magnetic Resonance Imaging, 28(5):621–628, 2010.[44] J. Kurhanewicz, D. Vigneron, P. Carroll, and F. Coakley. Multiparamet-ric Magnetic Resonance Imaging in prostate cancer: present and future.Current Opinion in Urology, 18(1):71–77, 2008.[45] D. L. Langer, T. H. van der Kwast, A. J. Evans, J. Trachtenberg, B. C.Wilson, and M. A. Haider. Prostate cancer detection: multi-parametricMRI with diffusion weighted imaging and dynamic contrast enhanced77BibliographyMRI. In Proceedings of the 17th Annual Meeting of ISMRM, Honolulu,page 782, 2009.[46] D. L. Langer, T. H. van der Kwast, A. J. Evans, J. Trachtenberg,B. C. Wilson, and M. A. Haider. Prostate cancer detection withmulti-parametric MRI: Logistic regression analysis of quantitative T2,diffusion-weighted imaging, and dynamic contrast-enhanced MRI. Jour-nal of Magnetic Resonance Imaging, 30(2):327–334, 2009.[47] D. L. Langer, T. H. van der Kwast, A. J. Evans, A. Plotkin, J. Tracht-enberg, B. C. Wilson, and M. A. Haider. Prostate Tissue Compositionand MR Measurements: Investigating the Relationships between ADC,T2, Ktrans, ve, and Corresponding Histologic Features. Radiology, 255(2):485–494, 2010.[48] D. Le Bihan, C. Poupon, A. Amadon, and F. Lethimonnier. Artifactsand pitfalls in diffusion MRI. Journal of Magnetic Resonance Imaging,24(3):478–488, 2006.[49] S. Lee, G. Wolberg, K. Y. Chwa, and S. Y. Shin. Image metamor-phosis with scattered feature constraints. Visualization and ComputerGraphics, IEEE Transactions on, 2(4):337–354, 1996.[50] S. Lee, G. Wolberg, and S. Y. Shin. Scattered data interpolationwith multilevel B-splines. Visualization and Computer Graphics, IEEETransactions on, 3(3):228–244, 1997.[51] G. P. Liney, L. W. Turnbull, and A. J. Knowles. In vivo magneticresonance spectroscopy and dynamic contrast enhanced imaging of theprostate gland. NMR in Biomedicine, 12(1):39–44, 1999.[52] D. G. Lowe. Distinctive image features from scale-invariant keypoints.International Journal of Computer Vision, 60(2):91–110, 2004.78Bibliography[53] M. Lowry, B. Zelhof, G. P. Liney, P. Gibbs, M. D. Pickles, and L. W.Turnbull. Analysis of prostate DCE-MRI: comparison of fast exchangelimit and fast exchange regimen pharmacokinetic models in the discrim-ination of malignant from normal tissue. Investigative Radiology, 44(9):577–584, 2009.[54] D. Lv, X. Guo, X. Wang, J. Zhang, and J. Fang. Computerized charac-terization of prostate cancer by fractal analysis in MR images. Journalof Magnetic Resonance Imaging, 30(1):161–168, 2009.[55] G. Manenti, M. Carlani, S. Mancino, V. Colangelo, M. Di Roma,E. Squillaci, and G. Simonetti. Diffusion tensor magnetic resonanceimaging of prostate cancer. Investigative Radiology, 42(6):412–419, 2007.[56] J. Martin, P. Hajek, L. Baker, V. Gylys-Morin, R. Fitzmorris-Glass,and R. Mattrey. Inflatable surface coil for MR imaging of the prostate.Radiology, 167(1):268–270, 1988.[57] D. Mattes, D. R. Haynor, H. Vesselle, T. K. Lewellen, and W. Eubank.PET-CT image registration in the chest using free-form deformations.Medical Imaging, IEEE Transactions on, 22(1):120–128, 2003.[58] P. S. Melki, R. V. Mulkern, L. P. Panych, and F. A. Jolesz. Comparingthe FAISE method with conventional dual-echo sequences. Journal ofMagnetic Resonance Imaging, 1(3):319–326, 1991.[59] R. Meng, S. D. Chang, E. C. Jones, S. L. Goldenberg, and P. Kozlowski.Comparison between population average and experimentally measuredarterial input function in predicting biopsy results in prostate cancer.Academic Radiology, 17(4):520–525, 2010.[60] M. Moradi, P. Abolmaesoumi, and P. Mousavi. Deformable registrationusing scale space keypoints. In Medical Imaging, pages 61442G–61442G.International Society for Optics and Photonics, 2006.79Bibliography[61] M. Moradi, S. E. Salcudean, S. D. Chang, E. C. Jones, N. Buchan, R. G.Casey, et al. Multiparametric MRI maps for detection and grading ofdominant prostate tumors. Journal of Magnetic Resonance Imaging, 35(6):1403–1413, 2012.[62] I. Ocak, M. Bernardo, G. Metzger, T. Barrett, P. Pinto, P. S. Albert,and P. L. Choyke. Dynamic contrast-enhanced MRI of prostate cancerat 3 T: a study of pharmacokinetic parameters. American Journal ofRoentgenology, 189(4):W192–W201, 2007.[63] G. J. Odowd, M. C. Miller, R. Orozco, and R. W. Veltri. Analysisof repeated biopsy results within 1 year after a noncancer diagnosis.Urology, 55(4):553–558, 2000.[64] A. Oto, A. Kayhan, Y. Jiang, M. Tretiakova, C. Yang, T. Antic, et al.Prostate Cancer: Differentiation of Central Gland Cancer from Be-nign Prostatic Hyperplasia by Using Diffusion-weighted and DynamicContrast-enhanced MR Imaging. Radiology, 257(3):715–723, 2010.[65] S. Ozer, D. L. Langer, X. Liu, M. A. Haider, T. H. van der Kwast, A. J.Evans, et al. Supervised and unsupervised methods for prostate cancersegmentation with multispectral MRI. Medical Physics, 37:1873–1883,2010.[66] A. R. Padhani, C. J. Gapinski, D. A. Macvicar, G. J. Parker, J. Suckling,P. B. Revell, et al. Dynamic Contrast Enhanced MRI of prostate cancer:correlation with morphology and tumour stage, histological grade andPSA. Clinical Radiology, 55(2):99–109, 2000.[67] G. Parker, A. Jackson, J. Waterton, and D. Buckley. Automated ArterialInput Function extraction for T1-weighted DCE-MRI. In Proceedings ofthe 11th Annual Meeting of ISMRM, Toronto, Canada, volume 11, page1264, 2003.80Bibliography[68] G. J. Parker, J. Suckling, S. F. Tanner, A. R. Padhani, P. B. Revell,J. E. Husband, and M. O. Leach. Probing tumor microvascularity bymeasurement, analysis and display of contrast agent uptake kinetics.Journal of Magnetic Resonance Imaging, 7(3):564–574, 1997.[69] G. J. Parker, C. Roberts, A. Macdonald, G. A. Buonaccorsi, S. Che-ung, D. L. Buckley, et al. Experimentally-derived functional form for apopulation-averaged high-temporal-resolution arterial input function forDynamic Contrast-Enhanced MRI. Magnetic Resonance in Medicine, 56(5):993–1000, 2006.[70] M. Perrotti, K.-R. Han, R. E. Epstein, E. C. Kennedy, F. Rabbani,K. Badani, et al. Prospective evaluation of endorectal magnetic reso-nance imaging to detect tumor foci in men with prior negative prostasticbiopsy: a pilot study. The Journal of urology, 162(4):1314–1317, 1999.[71] J. Platt. Probabilistic outputs for support vector machines and com-parisons to regularized likelihood methods. Advances in Large MarginClassifiers, 10(3):61–74, 1999.[72] P. Poon, R. McCallum, M. Henkelman, M. Bronskill, S. Sutcliffe,M. Jewett, et al. Magnetic Resonance Imaging of the prostate. Ra-diology, 154(1):143–149, 1985.[73] P. Preziosi, A. Orlacchio, G. Di Giambattista, P. Di Renzi, L. Bortolotti,A. Fabiano, et al. Enhancement patterns of prostate cancer in dynamicMRI. European Radiology, 13(5):925–930, 2003.[74] M. Rijpkema, J. H. Kaanders, F. Joosten, A. J. van der Kogel, andA. Heerschap. Method for quantitative mapping of dynamic MRI con-trast agent uptake in human tumors. Journal of Magnetic ResonanceImaging, 14(4):457–463, 2001.81Bibliography[75] C. Roberts, B. Issa, A. Stone, A. Jackson, J. C. Waterton, and G. J.Parker. Comparative study into the robustness of compartmental mod-eling and model-free analysis in DCE-MRI studies. Journal of MagneticResonance Imaging, 23(4):554–563, 2006.[76] O. Rouvie`re, A. Raudrant, R. Ecochard, C. Colin-Pangaud, C. Pasquiou,R. Bouvier, et al. Characterization of time-enhancement curves of be-nign and malignant prostate tissue at dynamic MR imaging. Europeanradiology, 13(5):931–942, 2003.[77] D. Rueckert, L. I. Sonoda, C. Hayes, D. L. Hill, M. O. Leach, and D. J.Hawkes. Nonrigid registration using free-form deformations: applicationto breast MR images. Medical Imaging, IEEE Transactions on, 18(8):712–721, 1999.[78] J. Salo, S. Rannikko, J. Ma¨kinen, and T. Lehtonen. Echogenic structureof prostatic cancer imaged on radical prostatectomy specimens. TheProstate, 10(1):1–9, 1987.[79] D. R. Salomao, S. D. Graham, and D. G. Bostwick. Microvascularinvasion in prostate cancer correlates with pathologic stage. Archives ofPathology & Laboratory Medicine, 119(11):1050–1054, 1995.[80] M. Scherr, M. Seitz, U. Mu¨ller-Lisse, M. Ingrisch, M. Reiser, andU. Mu¨ller-Lisse. MR-perfusion (MRP) and diffusion-weighted imaging(DWI) in prostate cancer: quantitative and model-based gadobenatedimeglumine MRP parameters in detection of prostate cancer. Euro-pean journal of radiology, 76(3):359–366, 2010.[81] M. D. Schnall, R. E. Lenkinski, H. M. Pollack, Y. Imai, and H. Y. Kres-sel. Prostate: MR imaging with an endorectal surface coil. Radiology,172(2):570–574, 1989.82Bibliography[82] F. H. Schro¨der, M. Roobol-Bouts, A. N. Vis, T. van der Kwast, andR. Kranse. Prostate-specific antigen-based early detection of prostatecancervalidation of screening without rectal examination. Urology, 57(1):83–90, 2001.[83] E. Steyerberg, M. Roobol, M. Kattan, T. Van der Kwast, H. De Koning,and F. Schro¨der. Prediction of indolent prostate cancer: validation andupdating of a prognostic nomogram. The Journal of Urology, 177(1):107–112, 2007.[84] T. Tamada, T. Sone, Y. Jo, S. Toshimitsu, T. Yamashita, A. Yamamoto,D. Tanimoto, and K. Ito. Apparent diffusion coefficient values in periph-eral and transition zones of the prostate: comparison between normaland malignant prostatic tissues and correlation with histologic grade.Journal of Magnetic Resonance Imaging, 28(3):720–726, 2008.[85] C. M. Tempany, X. Zhou, E. A. Zerhouni, M. D. Rifkin, L. E. Quint,C. W. Piccoli, et al. Staging of prostate cancer: results of radiology di-agnostic oncology group project comparison of three MR imaging tech-niques. Radiology, 192(1):47–54, 1994.[86] R. Tibshirani. Regression shrinkage and selection via the LASSO. Jour-nal of the Royal Statistical Society. Series B (Methodological), 58(1):267–288, 1996.[87] P. S. Tofts, B. Berkowitz, and M. D. Schnall. Quantitative analysis ofdynamic Gd-DTPA enhancement in breast tumors using a permeabilitymodel. Magnetic Resonance in Medicine, 33(4):564–568, 1995.[88] P. S. Tofts, G. Brix, D. L. Buckley, J. L. Evelhoch, E. Henderson, M. V.Knopp, et al. Estimating kinetic parameters from dynamic contrast-enhanced T1-weighted MRI of a diffusable tracer: standardized quanti-ties and symbols. Journal of Magnetic Resonance Imaging, 10(3):223–232, 1999.83Bibliography[89] C. G. van Niekerk, J. A. van der Laak, M. E. Bo¨rger, H.-J. Huisman,J. A. Witjes, J. O. Barentsz, and C. A. Hulsbergen-van de Kaa. Comput-erized whole slide quantification shows increased microvascular densityin pT2 prostate cancer as compared to normal prostate tissue. TheProstate, 69(1):62–69, 2009.[90] X. Z. Wang, B. Wang, Z. Q. Gao, J. G. Liu, Z. Q. Liu, Q. L. Niu, et al.Diffusion-weighted imaging of prostate cancer: Correlation between ap-parent diffusion coefficient values and tumor proliferation. Journal ofMagnetic Resonance Imaging, 29(6):1360–1366, 2009.[91] T. F. Wu, C. J. Lin, and R. C. Weng. Probability estimates for multi-class classification by pairwise coupling. Journal of Machine LearningResearch, 5(975-1005):4, 2004.[92] J. Xu, P. A. Humphrey, A. S. Kibel, A. Z. Snyder, V. R. Narra, J. J.Ackerman, and S. K. Song. Magnetic resonance diffusion characteristicsof histologically defined prostate cancer in humans. Magnetic Resonancein Medicine, 61(4):842–850, 2009.[93] T. Yoshizako, A. Wada, K. Uchida, S. Hara, M. Igawa, H. Kitagaki, andS. E. Maier. Apparent diffusion coefficient of line scan diffusion imagein normal prostate and prostate cancercomparison with single-shot echoplanner image. Magnetic resonance imaging, 29(1):106–110, 2011.[94] M. Zappa, S. Ciatto, R. Bonardi, and A. Mazzotta. Overdiagnosis ofprostate carcinoma by screening: an estimate based on the results of theFlorence Screening Pilot Study. Annals of Oncology, 9(12):1297–1300,1998.[95] B. Zelhof, M. Lowry, G. Rodrigues, S. Kraus, and L. Turnbull. De-scription of magnetic resonance imaging-derived enhancement variablesin pathologically confirmed prostate cancer and normal peripheral zoneregions. BJU international, 104(5):621–627, 2009.84