Ultrasound-based Approaches to Tissue Classification forBreast and Prostate Cancer DiagnosisbyNishant UniyalB.Sc. Electrical Engineering, Missouri University of Science & Technology, 2012B.Sc. Computer Engineering, Missouri University of Science & Technology, 2012A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFMasters of Applied ScienceinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Electrical and Computer Engineering)The University Of British Columbia(Vancouver)August 2014c© Nishant Uniyal, 2014AbstractUltrasound-based cancer diagnosis could improve the current breast and prostatecancer diagnosis methods. In this thesis, an ultrasound-based approach is evaluatedas a method for breast and prostate cancer diagnosis. Ultrasound RF time seriesalong with a comprehensive machine learning framework is used for accurate clas-sification of tissue samples. The RF time series method requires only a few secondsof raw ultrasound data with no need for additional instrumentation. The developedmethod produces cancer probability/likelihood maps which show the probability ofthe tissue under study being cancerous. These probability maps could provide ra-diologists with a real-time cancer diagnosis tool which could improve cancer yieldand significantly reduce the number of negative biopsies.To prove the utility of ultrasound RF time series as a tissue classification method,an in vivo breast lesion classification study of 22 subjects and an in vivo prostatebiopsy core classification study involving 18 subjects is presented in this thesis. Acomprehensive machine learning framework with a new semi-supervised learningtechnique for tissue classification is also presented in this work. An experimentalstudy to substantiate the ultrasound RF time series hypotheses by studying the ef-fects of ultrasound imaging parameters on animal tissue classification is also pre-sented. Using the ultrasound RF time series method and the developed machinelearning framework–we calculated the area under the receiver operating charac-teristics curve to be 85.6% for breast lesion classification and 91.5% for prostatetissue classification. Increasing the frame rate and the length of the time series,and decreasing the imaging depth we observed consistent improvement in tissueclassification results for the animal study.The results of this thesis suggest the potential of ultrasound RF time seriesiias a tissue classification method. Ultrasound RF time series along with otherultrasound-based methods could be a valuable and practical addition to the currentcancer diagnosis procedures. It has been shown here that a high level of accuracycan be attained using these tools which are non-invasive, inexpensive and readilyavailable to the clinician.iiiPrefaceA short version of Chapter 2 has been published in the proceedings of IEEE Ultra-sonics Symposium 2013 as ”A new approach to ultrasonic detection of malignantbreast tumors” 1. The article was co-authored by Hani Eskandari, Purang Abolmae-sumi, Samira Sojoudi, Paula Gordon, Linda Warren, Robert N. Rohling, SeptimiuE. Salcudean, and Mehdi Moradi. Chapter 2 has been submitted for publication inIEEE Transactions on Medical Imaging. A revised version was submitted in May2014. The author performed all the analysis on the data and developed the requiredalgorithms.The data collection for the breast study was performed during a study approvedby the Clinical Research Ethics Board at the University of British Columbia, Van-couver. The UBC CREB number of this study is H12-00889. Tim S.E. Salcudeanis the principal investigator for this study. This study was designed to evaluateultrasound-based tissue typing and vibro-elastography for breast lesion classifica-tion. The vibro-elastography technology was developed by Salcudean et al. HaniEskandari and Samira Sojoudi collected the data for this study.Chapter 3 is accepted for publication as a chapter in Lecture Notes in ComputerScience series and presentation at Medical Image Computing and Computer As-sisted Intervention (MICCAI) 2014 workshop on clinical image-based procedures,CLIP 2014. The report is co-authored by Farhad Imani, Amir Tahmasebi, HarshAgarwal, Shyam Bharat, Pingkun Yan, Jochen Kruecker, Jin Tae Kwak, Sheng Xu,Bradford Wood, Peter Pinto, Baris Turkbey, Peter Choyke, Purang Abolmaesumi,1[1] N. Uniyal, H. Eskandari, P. Abolmaesumi, S. Sojoudi, P. Gordon, L. Warren, R. N. Rohling,S. E. Salcudean, and M. Moradi, A new approach to ultrasonic detection of malignant breast tumors,in Ultrasonics Symposium (IUS), 2013 IEEE International, 2013, pp. 9699.ivParvin Mousavi, and Mehdi Moradi. Farhad Imani is also a lead co-author on thispaper.Data for the prostate study was collected at the National Institutes of HealthClinical Center (NIH-CC, Bethesda, MD). This work was a joint effort with NIH,Philips, and Queens University.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xGlossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviiiDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Thesis Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Organization of Thesis . . . . . . . . . . . . . . . . . . . . . . . 52 Ultrasound RF Time Series for Classification of Breast Lesions . . . 72.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . 102.2.2 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . 12vi2.2.3 Classification and Estimation of Cancer Likelihood . . . . 172.2.4 Evaluation of Classification . . . . . . . . . . . . . . . . 202.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212.3.1 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 212.3.2 Classification Results . . . . . . . . . . . . . . . . . . . . 212.4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 292.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 343 Ultrasound-based Prediction of Prostate Cancer in MRI-guided Biopsy 353.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 373.2.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . 373.2.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . 383.2.3 The Proposed Classification Framework . . . . . . . . . . 403.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413.4 Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . 443.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 484 Experiments to Provide Evidence on the Physical Basis of TissueClassification using RF Time Series . . . . . . . . . . . . . . . . . . 544.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544.2 Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . 554.2.1 Types of Tissue . . . . . . . . . . . . . . . . . . . . . . . 564.2.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . 574.2.3 Experiment Design . . . . . . . . . . . . . . . . . . . . . 584.2.4 Features . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.2.5 Classification and Cross-validation . . . . . . . . . . . . . 594.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604.3.1 Effects of Imaging Depth on Tissue Classification . . . . . 604.3.2 Effects of Frame Rate on Tissue Classification . . . . . . 614.3.3 Effects of Time Series Length on Tissue Classification . . 634.4 Discussions and Conclusions . . . . . . . . . . . . . . . . . . . . 634.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 65vii5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675.1 Summary of Contributions . . . . . . . . . . . . . . . . . . . . . 685.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71viiiList of TablesTable 2.1 Ultrasound RF time series, B-mode texture, and attenuation fea-tures calculated for classification of malignant and benign breastlesions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13Table 2.2 Sensitivity, specificity at a cut-off T hc=0.5, and area under ROCcurve. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Table 2.3 Classification results for the 22 subjects. . . . . . . . . . . . . 26Table 3.1 SVM and Random Forests cancer probabilities. . . . . . . . . 48Table 4.1 Parameters changed during each experiment. The parameters inbold were changed for the corresponding experiment. . . . . . 56Table 4.2 1 mm2 ROI size in the RF domain for different imaging depths. 59Table 4.3 Area under the ROC curve with changing imaging depth. . . . 60Table 4.4 Area under the ROC curve for different frame rates. . . . . . . 62Table 4.5 Area under the ROC curve for different time series length. . . . 63ixList of FiguresFigure 2.1 Overall classification framework. This figure shows a graphi-cal representation of the proposed approach. . . . . . . . . . . 10Figure 2.2 Figure showing the ROI selection. Each small rectangle in thelarger rectangle, inside the lesion, is an ROI and features werecalculated for each of the ROI. This figure does not correctlyrepresent the number of ROIs defined, it is for visual interpre-tation only. . . . . . . . . . . . . . . . . . . . . . . . . . . . 12Figure 2.3 The average power spectrum of the reflected RF signal forbreast tissue. . . . . . . . . . . . . . . . . . . . . . . . . . . 15Figure 2.4 Attenuation estimation by calculating the slope of the cali-brated spectrum. . . . . . . . . . . . . . . . . . . . . . . . . 16Figure 2.5 Box plots of the nine RF time series features. Table 2.1 pro-vides a description of the features. From the box plots, it isapparent that the best features selected by the RFE algorithmshow a difference in the class distribution. As can be seen inthe figure, ”Feature07” is the best feature that differentiates themalignant samples from the benign, and it was also selected asthe best feature by our feature selection method. . . . . . . . . 22Figure 2.6 Correlation plot of the 164 texture features extracted from theB-mode image. From the figure, it is observable that most ofthe features are highly correlated. High positive correlationsuggests the need for feature selection prior to classification. . 23Figure 2.7 ROC curve for SVM and Random Forests using three features. 24xFigure 2.8 Classification results for the 22 subjects. As can be seen in thefigure, five out of seven malignant subjects were correctly clas-sified and 13 out of 15 benign subjects were correctly classified. 25Figure 2.9 The malignancy map created by plotting the value cancer like-lihood (Lc) of the ROIs overlaid on the B-mode image. Theimages on the top are from malignant subjects, and the oneson the bottom are benign subjects. . . . . . . . . . . . . . . . 25Figure 2.10 The malignancy maps for the subjects that the classifiers didnot correctly label. On the top are two incorrectly classifiedmalignant subjects (Subjects 3 and 7), and the ones on the bot-tom row are the incorrectly classified benign subjects (Subjects10 and 16). . . . . . . . . . . . . . . . . . . . . . . . . . . . 27Figure 2.11 Scatter plot showing the ROI samples from the 22 subjects us-ing features 2 and 3. Red and blue markers represent malignantand benign ROI samples from 22 subjects. . . . . . . . . . . . 28Figure 2.12 SVM decision function plot for the malignant subjects. Thered region is the decision boundary for the malignant ROI sam-ples and the cyan region is the decision boundary for the be-nign ROI samples. As can be seen in the figure, malignantsubjects, 3 and 7, lie in the benign region and are thereforemisclassified by the classification algorithms. See Figure 2.11for distribution of the ROI samples. . . . . . . . . . . . . . . 29Figure 2.13 Random Forests decision function plot for the malignant sub-jects. The red region is the decision boundary for the malignantROI samples and the cyan region is the decision boundary forthe benign ROI samples. As can be seen in the figure, malig-nant subjects, 3 and 7, lie in the benign region and are thereforemisclassified by the classification algorithms. See Figure 2.11for distribution of the ROI samples. . . . . . . . . . . . . . . 30xiFigure 2.14 SVM decision function plot for the benign subjects. The redregion is the decision boundary for the malignant ROI samplesand the cyan region is the decision boundary for the benignROI samples. As can be seen in the figure, benign subjects, 10and 16, lie in the malignant region and are therefore misclas-sified by the classification algorithm. Subject 10 is an outlierin the dataset and significantly effects the classification results.See Figure 2.11 for distribution of the ROI samples. . . . . . . 31Figure 2.15 Random Forests decision function plot for the benign subjects.The red region is the decision boundary for the malignant ROIsamples and the cyan region is the decision boundary for thebenign ROI samples. As can be seen in the figure, benign sub-jects, 10 and 16, lie in the malignant region and are thereforemisclassified by the classification algorithm. Subject 10 is anoutlier in the dataset and significantly effects the classificationresults. See Figure 2.11 for distribution of the ROI samples. . 32Figure 3.1 Philips UroNav platform (left) and a Philips iU22 ultrasoundmachine (right). . . . . . . . . . . . . . . . . . . . . . . . . . 38Figure 3.2 B-mode image showing the ROI selection. ROIs were selectedalong the biopsy needle trajectory as shown in the figure. . . . 39Figure 3.3 An overview of the classification framework employed in thiswork. Clustering of the data is performed prior to classificationto eliminate the affect of outliers and improve the classificationaccuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40Figure 3.4 Boxplots of features. . . . . . . . . . . . . . . . . . . . . . . 42Figure 3.5 Correlation plots of the RF time series features. As can beseen in the figure, all the time series features are positivelycorrelated, except for features 6 and 9. Due to high correla-tion among the features, we perform feature selection beforeclassification. . . . . . . . . . . . . . . . . . . . . . . . . . . 43Figure 3.6 Histograms of the features. The best selected features, features3 and 4, are highlited in red. . . . . . . . . . . . . . . . . . . 44xiiFigure 3.7 Clustering performed on all 360 training samples. . . . . . . . 45Figure 3.8 ROC curves for SVM and Random Forests (each performedafter clustering). . . . . . . . . . . . . . . . . . . . . . . . . 46Figure 3.9 Classification results for the 18 bipsy cores. As can be seen inthe figure, all biopsy cores were correctly classified. . . . . . 47Figure 3.10 Cancer likelihood colormaps of the 18 biopsy cores from 14subjects with leave-one-subject-out cross validation using thebest two RF time series features. Clustering and SVM classifi-cation is used. . . . . . . . . . . . . . . . . . . . . . . . . . . 47Figure 3.11 Scatter plot showing the ROI samples from the 18 biopsy coresusing features 3 and 4. Red and blue markers represent malig-nant and benign ROI samples from 18 cores. . . . . . . . . . 49Figure 3.12 SVM decision function plots for the malignant cores. The de-cision plots shown here are without clustering. The red regionis the decision boundary for the malignant ROI samples andthe cyan region is the decision boundary for the benign ROIsamples. It is apparent in the above figure, some ROI samplesfrom cores, 3, 8, 9, and 11 are in the benign region and with-out clustering would have been misclassified. These decisionfunctions (plotted without clustering) show the utility of clus-tering prior to classification. See Figure 3.11 for distributionof the ROI samples. . . . . . . . . . . . . . . . . . . . . . . . 50Figure 3.13 Random Forests decision function plots for the malignant cores.The decision plots shown here are without clustering. The redregion is the decision boundary for the malignant ROI samplesand the cyan region is the decision boundary for the benignROI samples. It is apparent in the above figure, some ROIsamples from cores, 3, 8, 9, and 11 are in the benign regionand without clustering would have been misclassified. Thesedecision functions (plotted without clustering) show the utilityof clustering prior to classification. See Figure 3.11 for distri-bution of the ROI samples. . . . . . . . . . . . . . . . . . . . 51xiiiFigure 3.14 SVM decision function plots for the benign cores. The deci-sion plots shown here are without clustering. The red regionis the decision boundary for the malignant ROI samples andthe cyan region is the decision boundary for the benign ROIsamples. As can be seen in the above figure, most of the be-nign ROI samples lie in the benign region (cyan). Some ROIsamples from core 6 are in the malignant region and hence theclassification accuracy for core 6 is lower than the other benigncores. See Figure 3.11 for distribution of the ROI samples. . . 52Figure 3.15 Random Forests decision function plots for the benign cores.The decision plots shown here are without clustering. The redregion is the decision boundary for the malignant ROI samplesand the cyan region is the decision boundary for the benignROI samples. As can be seen in the above figure, most of thebenign ROI samples lie in the benign region (cyan). Some ROIsamples from core 6 are in the malignant region and hence theclassification accuracy for core 6 is lower than the other benigncores. See Figure 3.11 for distribution of the ROI samples. . . 53Figure 4.1 Experiment setup. The two animal tissues are Chicken breaston the left and bovine on the right surrounded by an ultrasoundabsorbing pad. . . . . . . . . . . . . . . . . . . . . . . . . . 57Figure 4.2 B-mode image of the two animal tissues. Bovine on the leftand chicken (right). . . . . . . . . . . . . . . . . . . . . . . . 58Figure 4.3 The effect of imaging depth on the classification accuracy ofthe two tissue types, bovine and chicken. As can be seen inthe figure, increasing the imaging depth decreased the classi-fication accuracy. In other words, using the RF time seriesmethod, targets that are closer to the ultrasound transducer canbe classified more accurately than the targets that are fartheraway from the transducer. The results are consistent with theprevious studies. . . . . . . . . . . . . . . . . . . . . . . . . 61xivFigure 4.4 The effect of increasing frame rate on the classification accu-racy of the two tissue types, bovine and chicken. Increasingthe frame rate significantly improved the classification results.It is apparent from the above figure, increasing the frame ratefrom 40 to 71 frames per second improved the area under theROC curve from 77% to 87%. Previous studies have shownthat increasing the frame rate improved the classification per-formance, therefore our results are consistent with the previousstudies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62Figure 4.5 The effect of the length of the time series on the classificationaccuracy of the two tissue types, bovine and chicken. As canbe seen in the above figure, increasing the number of analysedRF frames, for the RF time series feature calculation, improvedthe classification performance. These results show, the tissuetyping information improved as more RF frames are analysed.These observations are consistent with the previous publishedliterature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64xvGlossary3D three-dimensional2D two-dimensionalRF radio-frequencySVM support vector machinesMRI magnetic resonance imagingCT computer tomographyRFE recursive feature eliminationDCE dynamic contrast enhancedDWI diffusion weighted imagingTRUS transrectal ultrasoundCI confidence intervalEM electromagneticROI region of interestROC receiver operating characteristicTGC time gain controlGLCM gray-level co-occurrence matrixxviRBF radial basis functionDFT discrete Fourier transformMCF mean center frequencyxviiAcknowledgmentsI would first like to thank my supervisor, Dr. Mehdi Moradi for the support,guidance, and the encouragement throughout my post graduate studies. WithoutMehdi’s motivation this thesis would not have been possible. Dr. Purang Abolmae-sumi for his support and finding time to proofread the manuscripts and providinginvaluable comments. Dr. Robert Rohling and Dr. Septimiu Salcudean for theirhelp with the manuscripts and sharing their expertise.A special thanks to Hani Eskandari for providing the data for the breast study,helping with the manuscript, and for the guidance throughout my learning phase.Farhad Imani and Dr. Parvin Mousavi for their help on the prostate study. BoZhuang for helping with the data collection for the animal tissue experiments. AmirTahmasebi and Shyam Bharat for their help while I was at Philips.Thank you to all my friends and colleagues in the Robotics and Controls Labat UBC. Thank you for a friendly lab environment and for all the social gatherings.I will miss our Friday lunches and the lab events.Last but not the least, I would like to thank my mother for her support andencouragement throughout my career. Without her I wouldn’t have been able toachieve this milestone. Thank you to my sister who has always encouraged me topursue my dreams.Finally, I would like to thank the funding sources for my project and research:The Natural Sciences and Engineering Research Council of Canada, Canadian In-stitutes of Health Research, and the Canadian Foundation for Innovation.xviiiDedicationTo my motherxixChapter 1IntroductionIf I have seen farther it is by standing on the shoulders of Giants.— Sir Isaac Newton (1855)1.1 MotivationAfter lung cancer, breast and prostate cancer are the most common causes of cancerdeaths in North America [1]. Among females, breast cancer is the leading cause ofdeath and the most frequently diagnosed cancer in the world [2]. Early detectionand advanced diagnosis play a vital role in reducing the number of fatalities due tobreast and prostate cancer and medical imaging is a promising assessment tool forthis purpose.Imaging modalities including, X-ray or computer tomography (CT), magneticresonance imaging (MRI), and ultrasound have been used for breast and prostatecancer screening and diagnosis. Ultrasound is the most versatile and widely usedimaging modality. Unlike X-ray, ultrasound does not expose the human body toionizing radiation. Ultrasound is less expensive, more portable, and more widelyavailable than MRI.Mammography, X-ray of the breast, is the primary screening modality forbreast cancer. Subjects with a suspicious lesion in mammography undergo an ul-trasound examination before an ultrasound-guided biopsy is performed. However,almost 80% of the biopsies carried out turn out to be negative for cancer [3, 4].1Recently, a 25-year study of 89,835 women in Canada, aged 40-59, found thatmammography screening does not reduce mortality from breast cancer and resultsin over-diagnosis of breast cancer [5]. Mammography alone is limited in its abilityto detect cancer in dense breasts [6, 7]. Ultrasound, along with mammography,has been proven to improve the sensitivity in palpable breast masses [8–10] butdue to its low sensitivity towards non-palpable and non-cystic breast lesions, it isnot used as a primary screening modality. Ultrasound has also been found to beuseful in classification of benign and malignant breast lesions [4, 11]. However,low contrast, high speckle brightness mode (B-mode) images, and operator vari-ability make ultrasound based diagnosis more difficult [11]. To address the limita-tion of ultrasound B-mode breast tissue classification several research groups haveproposed alternative methods for breast tissue classification [12–17]. A thoroughliterature review of ultrasound-based breast lesion classification methods can befound in Section 2.1.Similarly, MRI has been used for prostate cancer diagnosis to improve highgrade prostate cancer yield [18–21]. However, MRI-based prostate cancer di-agnosis is difficult, expensive, and time-consuming [22]. MRI in this contextrefers to multi-parametric MRI (mpMRI) which combines dynamic contrast en-hanced (DCE) and diffusion weighted imaging (DWI) [23]. Fusion of ultrasoundand MRI has been used to improve prostate cancer detection by enabling targetingof the cancer foci pre-determined in mpMRI during transrectal ultrasound (TRUS)-guided prostate biopsy [18]. Biopsy core locations determined by a radiologistthrough examination of mpMRI data and images are translated to patient coor-dinates using pre-procedure three-dimensional (3D) TRUS and its registration toMRI [24]. 3D TRUS to MRI registration requires either sophisticated mechanicalsystems [25] to guide the biopsy needles or, if performed by software only [26],does not fully account for patient motion or organ deformation occurring duringbiopsy. Alternatively, other ultrasound-based prostate cancer diagnosis techniquessuch as radio-frequency (RF) data analysis [27], elastography [28, 29], and Dopplerimaging [30] have been reported [31]. However, these technologies, individually,have not entirely succeeded in accurate identification of high grade cancer.In recent works, ultrasound RF time series, comprising a sequence of ultra-sound RF frames captured in time from a stationary tissue location, has been pro-2posed as a new tissue classification method [32]. In this method of analysis, thetissue typing parameters are extracted from the temporal changes of the signal, asopposed to the classical method of spectral analysis on spatial segments of the RFsignals [33]. Ultrasound RF time series method for tissue classification is com-pletely non invasive and does not require specialized equipment. As such, it canbe a valuable and practical addition to other ultrasound-based methods for tissueclassification such as elastography or single frame RF analysis. The next paragraphprovides a history and literature review of ultrasound RF time series.Ultrasound RF time series was first proposed back in 2006 when it was usedfor detection of prostate cancer [34]. Moradi et al. published a study on fractalanalysis of ultrasound RF time series signals for prostate cancer diagnosis. Theirwork presented the use of a fractal dimension feature and B-mode texture featuresfor prostate tissue classification [34]. After that, Moradi et al. [35] described astudy in which they performed discrete Fourier transform (DFT) of ultrasound RFtime series signals and calculated six features. A neural network classifier wasthen employed to classify region of interest (ROI) samples described by the sixDFT features. Their work reported 91% mean accuracy and a mean sensitivity andspecificity of 92% and 90%. In the same year, Moradi et al. [36] reported theuse of high frequency RF time series data for classification of animal tissue. Theycalculated the fractal dimension features from the high frequency RF time seriessignal and used a Bayesian classifier to estimate posterior class probabilities foranimal tissue classification. Their work reported tissue classification accuracy ofupto 98%. In another study, Moradi et al. [32] used ultrasound RF time seriesfor classifying prostate cancer in 35 ex vivo specimens. Using a leave-one-patient-out cross-validation scheme, a support vector machines (SVM) classifier, and RFtime series features they reported an area under the receiver operating characteris-tic (ROC) curve of 0.82. In 2010, Moradi et al. [37] used ultrasound RF time seriesfeatures to differentiate animal tissue samples. In their report, they also studied theeffects of ultrasound imaging parameters, such as transmit power and frame rate,on tissue classification. Their work showed that increased energy delivered throughpower and frame rate improves the tissue classification results. Their results alsoreported high classification accuracy of 95% at 6.6 MHz and 98% at 55 MHz. In2011, Imani et al. [38] extracted novel features, mean center frequency (MCF) val-3ues and slope and intercept fitted to MCF values, from ultrasound RF time seriessignal to differentiate in vitro animal tissue samples . Using the MCF features andan SVM classifier their study reported 99-100% accuracy in differentiating bovinemuscle, bovine liver, and chicken breast. In 2013, various in vivo studies reportingthe use of RF time series were published. Moradi et al. [39] published the first invivo paper employing RF time series features and SVM to classify prostate tissuesamples from six subjects. In their work, they reported an average area under theROC curve of 0.76. Imani et al. [40] in their report on in vivo prostate tissue clas-sification proposed novel wavelet features extracted from RF time series signal.Their work, reported an average area under the ROC curve of 0.83 on data fromseven subjects. Imani et al. also showed that their proposed features significantlyoutperformed the previous proposed six DFT RF time series features. RF time se-ries method was also successfully used by Imani et al. [41] for differentiation ofablated and non-ablated tissue. In this work, ultrasound RF time series has beenused for the first time to classify breast lesions. A new semi-supervised machinelearning framework is developed to classify MRI-targeted prostate biopsy cores.An animal tissue study is completed to provide experimental evidence on the phys-ical basis of tissue typing using RF time series and prove that is can be used as atissue typing method.In the previous paragraph, we discussed various studies that showed that ultra-sound RF time series method can be used for tissue classification. While ultrasoundRF time series method has yielded consistent results, the source of tissue classifica-tion using this method is still mostly unknown. This suggests a need for additionalexperimental studies to address the physical basis of the tissue typing informa-tion extracted from ultrasound RF time series. Several hypotheses describing thephysical phenomenon exists and have been explored in different studies. Daoud etal. have shown experimentally and in simulations, that the temperature increasein controlled irradiation of tissue with RF time series can partly explain the phe-nomenon. Also, Moradi et al. [37] and Imani et al. [41] reported that increasingthe energy delivered through transmit power and frame rate improves the tissueclassification accuracy in phantom and animal studies. Given these observations,micro-vibrations of the tissue microstructure caused by acoustic radiation forceis another likely phenomenon. The acoustic radiation force is related to both the4acoustic energy and to the attenuation and scattering properties of the tissue, whichare different for different pathological tissue types. To substantiate the tissue clas-sification power of ultrasound RF time series and build upon previous hypotheses,an animal tissue experimental study is presented in this thesis.1.2 Thesis ObjectivesThe overall objective of this thesis was to develop a tissue classification frameworkand test it on breast and prostate dataset.1. To develop a breast lesion classification framework based on features ex-tracted from ultrasound RF spectral data, B-mode image texture data, and ultra-sound RF time series data: a machine learning package was developed to classifylesions described by the extracted features. The outcome of this study could poten-tially improve the current breast cancer diagnosis methods and provide radiologistswith a readily available diagnosis tool.2. To augment the current prostate cancer diagnosis by developing an innova-tive machine learning framework to classify MRI-targeted prostate biopsy samples:in this study, previously proposed ultrasound RF time series features were used ina new machine learning package to classify prostate biopsy samples with high ac-curacy. This work could potentially lead to the fusion of the proposed tissue classi-fication framework with the Philips UroNav platform to complement MRI-targetedTRUS guided prostate biopsy and provide real-time diagnosis of prostate cancer.3. To substantiate the RF time series hypotheses by studying the effects ofultrasound imaging parameters such as imaging depth, frame rate, and time serieslength, on tissue classification through experiments on animal tissue: this part ofthe work could provide more evidence on the physical basis of tissue classificationusing ultrasound RF time series method.1.3 Organization of ThesisThe outline of the thesis is as follows:Chapter 2 describes a study on breast lesion classification using ultrasound RFtime series features and B-mode texture features along with a machine learningframework. This is the first report on using the RF time series analysis in the5context of breast lesions classification.Chapter 3 is based on prediction of prostate cancer in MRI-targeted biopsy.Ultrasound RF time series features are used within an innovative semi-supervisedclassification framework. In this chapter, a new semi-supervised machine learningframework is presented that improves prostate tissue classification.Chapter 4 presents the results of experiments conducted on animal tissue tostudy the effects of ultrasound imaging parameters, such as imaging depth, framerate, and time series length, on tissue classification. This chapter addresses theneed for experimental evidence on the RF time series hypothesis.Lastly, Chapter 5 describes the conclusions, thesis contributions, and futurework in this field.6Chapter 2Ultrasound RF Time Series forClassification of Breast Lesions2.1 IntroductionAccording to the statistics published by the American Cancer Society in 2013, itwas estimated that breast cancer is one of the most common types of cancer inwomen accounting for 29% (232,340) of all cancer cases [42]. The World HealthOrganization estimated that almost 1.38 million women worldwide are diagnosedwith breast cancer annually. That accounts for 23% of all cancer cases [43]. Earlydetection and better diagnosis methods play a significant role in reducing the num-ber of fatalities due to breast cancer.Based on mammography images, the American College of Radiology has de-veloped a method called BI-RADS (Breast Imaging-Reporting and Data System)for deciding whether biopsy of an identified suspicious lesion is indicated. Biopsyis the gold standard for breast cancer diagnosis; however it is an expensive, dis-comforting, and invasive procedure. Almost 80% of the biopsies carried out basedon BI-RADS score turn out to be benign [3]. Hence, there exists a need for reduc-ing the number of unnecessary breast biopsies and augment the current diagnosismethods.Mammography alone is limited in its ability to detect cancer. Studies report73% sensitivity for all breast types and only 48% for dense breasts [8]. Almost750% of women under the age of 50 have dense breasts [44] and could be potentiallydeemed undiagnosed after mammography screening [7, 45]. Ultrasound, alongwith mammography, has been proven to improve the sensitivity in palpable breastmasses [8–10], but due to its low sensitivity towards non-palpable and non-cysticbreast lesions, it is not used as a primary screening modality. Ultrasound has alsobeen found to be useful in classification of benign and malignant breast lesions[4, 11]. However, its drawback lies in low contrast, high speckle B-mode images,and operator variability [11]. To address these limitations, several research groupsare working towards ultrasound-based computer-aided diagnosis (CAD) for breastcancer diagnosis.For example, Zheng et al. [46] used sonographic texture features along withself-organizing maps for classification of breast abnormalities. Chen et al. [47]also used neural networks and texture information extracted from ultrasound im-ages to classify breast nodules. Giger et al. [48] calculated features related to lesionmargin, shape, homogeneity (texture), and posterior acoustic attenuation patternsin ultrasound images and used linear discriminant analysis to classify breast le-sions. Lefebvre et al. [49] used texture and morphometric parameters with lineardiscriminant analysis and leave-one-out cross-validation to classify breast lesions.Donohue et al. [50] described a breast lesion classification method using textureand generalized spectrum features along with linear and quadratic discriminantanalysis. Shankar et al. investigated the Nakagami distribution [51] and non-Rayleigh statistics [52] of the backscattered envelope for classification of breastlesions. Chang et al. [53, 54] and Huang et al. [55] used the same texture analysisapproach but employed support vector machines to classify breast lesions. Drukkeret al. [56, 57] performed analysis of posterior acoustic shadowing for breast lesiondetection. Joo et al. [58] employed an artificial neural network to detect breastlesions based on five morphological features representing the shape, edge charac-teristics, and darkness of a lesion. Chen et al. [59] described an approach basedon fractal parameter computation from ultrasound images and the k-means classi-fication algorithm. Alam et al. [13] reported the performance of logistic regressionin classification of breast lesions based on quantitative acoustic parameters calcu-lated using the spectrum analysis of ultrasound RF echo signals and morphometricfeatures related to the lesion shape. In a more recent work, Gomez et al. [60] inves-8tigated the effectiveness of co-occurrence texture statistics calculated at differentquantization levels, as a method for classification of breast images. Yang et al. [61]reported another method for ultrasound image diagnosis using gray-scale invariantfeatures extracted via multi-resolution ranklet transform along with support vectormachines. Tan et al. [17] developed a CAD system to detect cancer in 3-D breastultrasound. A detailed review of the literature in the area of ultrasound-based CADfor breast cancer is reported in [12].Ultrasound strain imaging [14, 62], Acoustic Radiation Force Impulse (ARFI)imaging [16], and supersonic shear-wave imaging [15, 63–65] have also been usedfor breast lesion characterization. These methodologies require the use of excita-tion mechanisms or specialized equipment.As introduced in Chapter 1, ultrasound RF time series analysis has been pro-posed as a novel tissue classification method by our research group. The RF timeseries method is based on a temporal analysis of beamformed RF signals. Thismethod does not require additional equipment and the region of interest selection isas simple as drawing a rectangular box inside the lesion. The task of size selectionand placement of the rectangular box is much less subjective compared to manualcontouring of the lesions. As such, it can be a valuable and practical addition tothe previously reported ultrasound-based methods for breast lesion classification.This is the first report of using the RF time series analysis in the context oflesion classification in breast images. Preliminary results have been previouslypresented in [66]. For this study, we use quantitative analysis of spectral and frac-tal parameters extracted from RF time series. We report the use of both SVM andRandom Forests classification methods with RF time series features and demon-strate accurate malignancy maps that can be used for decision support in biopsyrecommendation.2.2 MethodsWe describe the use of a machine learning framework to assign labels to eachsample and also estimate the cancer likelihood (for SVM) and probability (in caseof Random Forests). The estimated cancer likelihood is then used to generate amalignancy map. Figure 2.1 shows a graphical representation of this approach.9Figure 2.1: Overall classification framework. This figure shows a graphicalrepresentation of the proposed approach.2.2.1 Data CollectionAn RF time series is formed by the sequence of RF echoes received from onelocation in the tissue over time. To acquire the RF time series data, the ultrasoundprobe and the tissue remain fixed and frames of RF signals are acquired. In thismethod of analysis, the tissue typing parameters are extracted from the temporalchanges of the signal, as opposed to the classical method of spectral analysis onspatial segments of the RF signals [33].The data collection for this work was performed during a study approved by theClinical Research Ethics Board at the University of British Columbia. This studywas designed to evaluate ultrasound-based tissue typing and vibro-elastography forbreast lesion classification. The vibro-elastography technology was developed bySalcudean et al., and the preliminary results were reported in [67]. Subjects re-ferred to ultrasound-guided biopsy, based on mammography screening, were con-10sented for collection of vibro-elastography and RF data during biopsy. The studywas conducted between September 2012 and January 2013. Data was obtainedon a SonixTouch ultrasound machine (Ultrasonix Medical Corp., Richmond, BC,Canada). The research platform provided by the manufacturer enabled acquisitionof beamformed RF signals in real time.For every subject, the sonographer first performed a preliminary ultrasoundscan to find the suspected lesion. Once the lesion was located, the sonographerwould hold her hand steady for 5 seconds while a computer program stored the RFdata into the memory and consequently saved it in a file. Imaging was performedwith an L14-5/38 ultrasound transducer at a center frequency of 10 MHz and adepth of 4 cm. Each RF line was sampled at 40 MHz and a total of 128 scan-lines were acquired for each RF frame. With these image settings, RF data ata frame-rate of 98 frames per second, for five seconds, was obtained. The datacollection of each subject was followed by a routine ultrasound exam and a coreneedle biopsy of the lesion under ultrasound guidance by the physician. In total, 35subjects were imaged. However, in five subjects the radiologist decided that biopsywas unnecessary. Also, in eight subjects where biopsies were performed, the exactlocation of the biopsy sample could not be identified within the ultrasound image.As a result, we used the data corresponding to 22 subjects for whom the biopsyresult was available and could be mapped to a specific lesion. Biopsy results forthese subjects showed seven malignant lesions, all of the invasive ductal carcinomatype and 15 benign subjects, mostly of fibroadenoma type.To use the RF time series method, patient motion has to be minimized or com-pensated. In the current study, we used only the first 128 frames of the data, whichaccounts for 1.3 seconds of acquisition, to form the RF time series. Our motionestimation using the method described in [68] showed that when the time seriesis limited to 128 frames, the displacement caused by patient motion is on averagesmaller than the resolution of the RF samples. This means that we could, on aver-age, assume that series of RF samples obtained at a certain image coordinate were,in fact, from the same physical location in the tissue.112.2.2 FeaturesThe biopsied lesions were divided into 1 mm2 regions of interest (ROIs). The tissuetyping features were extracted from these ROIs. In the RF domain, this ROI sizewas equivalent to 3×52 samples each forming a time series. In the interpolatedB-mode image the ROI size of 1 mm2 was equivalent to 26×26 pixels. The RFframe size was 128×2080 and the B-mode image resolution was 988×1040. Notethat from the 22 lesions in the dataset, a total of 863 ROIs were extracted. Amongthese, 241 were malignant. From each subject, we extracted multiple ROI and thenumber of ROIs depends on the size of the lesion.Figure 2.2: Figure showing the ROI selection. Each small rectangle in thelarger rectangle, inside the lesion, is an ROI and features were calculatedfor each of the ROI. This figure does not correctly represent the numberof ROIs defined, it is for visual interpretation only.In addition to the RF time series features, we also calculated and used thevalues of texture features extracted from the first B-mode image and spectral RF12features extracted from the first RF frame. A similar approach to B-mode textureprofiling for tissue typing is reported in prior work [47–50, 60, 61, 69–71]. Acomplete list of the studied parameters can be found in Table 2.1.Table 2.1: Ultrasound RF time series, B-mode texture, and attenuation featurescalculated for classification of malignant and benign breast lesions# Feature Origin1 RF time series first quadrant RF time series2 RF time series second quadrant RF time series3 RF time series third quadrant RF time series4 RF time series fourth quadrant RF time series5 Intercept of regression line fitted to norm. spect. RF time series6 Slope of regression line fitted to norm. spect. RF time series7 Higuchi fractal dimension RF time series8 Intercept of line fitted to calibrated spectrum Single-frame RF9 Slope of line fitted to calibrated spectrum Single-frame RF10 Texture: Mean B-mode11 Texture: Standard Deviation B-mode12 Texture: Skewness B-mode13 Texture: Kurtosis B-mode14-17 Correlation:Fd(θ) =∑i, j(i−µi)( j−µ j)p(i, j)σiσ jB-mode18-21 Energy:Fd(θ) =∑i, jp(i, j)2 B-mode22-25 Contrast:Fd(θ) =∑i, j|i− j|2 p(i, j) B-mode26-29 Homogeneity:Fd(θ) =∑i, jp(i, j)1+ |i− j|B-mode† Features 14-29 are calculated for each d ∈ (1, 2,..,10) resulting in 160 (16×10) feature values.θ is the orientation angle for GLCM calculation (θ = 0o,45o,90o,135o).d is the pixel distance for GLCM calculation.p(i, j) is the (i, j)th entry in the co-occurrence probability matrix.µi =∑i, ji · p(i, j) and µ j =∑i, jj · p(i, j)σ2i =∑i, j(i−µi)2 · p(i, j) and σ2j =∑i, j( j−µ j)2 · p(i, j)RF time series features: A variety of methods have been used for tissue classifi-13cation using RF time series. We use the method originally described in [35] whichproposes summarizing the power spectrum of the RF time series in six parameters.The frequency spectrum was estimated by calculating the FFT-based periodogramof the Hamming windowed time series. This estimated spectrum was divided intofour frequency bands and each averaged to deliver a feature. In other words, thefirst four features (Features 1-4) were the average of the frequency spectrum in[0,pi/4), [pi/4,pi/2), [pi/2,3pi/4), [3pi/4,pi) frequency bands in the discrete frequencydomain. Note that the sampling rate here is equivalent to the frame rate of the ul-trasound machine. Two other spectral features were the intercept (Feature 5) andthe slope (Feature 6) of a regression line fitted to the magnitude of the spectrumversus normalized frequency.Feature 7 was the average fractal dimension (FD) of RF time series in a regionof interest. In this context, fractal dimension is a measure of the non-linear com-plexity of the signal. For calculation of the FD, we used the algorithm proposed byHiguchi [72] which decomposes the signal into different scales and evaluates thesignal complexity. We used the Higuchi algorithm with 16 levels of decompositionfor the time series of length 128. The use of this feature in ultrasound-based tissuetyping was originally proposed in [34].Single-frame RF spectral features: The tissue attenuation coefficient and RFspectral parameters that are related to attenuation have been used for tissue typing.Similar to texture features, these features were also calculated using the first RFframe. One common method for evaluating the frequency-dependent attenuationof ultrasound in tissue is the spectral analysis of segments of echoed RF signals,after calibration to remove the effects of the imaging system. Researchers at theLizzi Center for Biomedical Engineering at Riverside Research have shown that theparameters of a linear model fitted to the calibrated power spectrum of an RF seg-ment contain valuable tissue classification information [73, 74]. To estimate theseparameters, the averaged tissue power spectrum within an ROI had to be calibratedto account for the ultrasound machine transfer function. The calibration spectrumwas acquired from the surface of a flat glass plate in a water bath at the transducerfocal zone, with minimum amplifier gain and flat time gain control (TGC). Follow-ing [73], we subtracted the logarithm of the averaged spectrum of the glass fromthe logarithm of the spectrum of the tissue. For our data, we found that the useful14part of the spectrum was at the range of 1-10 MHz and this is apparent in Figure2.3. Outside of this range, the power drops to less than half of the peak value.Therefore, we fitted the regression line to the data in this range. As can be seen in,Figure 2.4 the calibrated spectrum shows a quasi-linear behaviour. Features 8 and9 are the intercept and slope of the linear model fitted to the normalized frequencyspectrum.0 5 10 15 20−50−45−40−35−30−25−20−15−10−5Frequency (MHz)Power (dB) Figure 2.3: The average power spectrum of the reflected RF signal for breasttissue.B-mode Texture features: Features 10–173 are 164 texture parameters calcu-lated using the B-mode image reconstructed from the first RF frame. The RFframe was converted into a B-mode image by taking the Hilbert transform andthen resized to 988×1040 to match the physical dimension (3.8×4 cm) of the im-age. The first four parameters were simply the mean, standard deviation, skewnessand kurtosis of the pixel intensities in an ROI and the other 160 parameters werethe correlation, energy, contrast, and homogeneity calculated from the gray-level151 2 3 4 5 6 7 8 9 10−25−20−15−10Frequency (MHz)Amplitude (dB)Linear model fitted to a normalized spectrum Calibrated/Normalized spectrum Fitted linear modelFigure 2.4: Attenuation estimation by calculating the slope of the calibratedspectrum.co-occurrence matrix (GLCM). The GLCM was created by calculating how often apixel with a grayscale intensity i occurs adjacent to a pixel with the value j. For thepurpose of this study, we calculated the co-occurrence matrices at a distance of tenpixel (d = 1,2, ...,10), for directions of θ = 0o, θ = 45o, θ = 90o, and θ = 135o.For every one of the 40 combinations of d and θ values a GLCM was calculated.Four texture statistics are calculated from each GLCM and the definition of thesefeatures is provided in Table 2.1.We also calculated the single-frame RF spectral features and B-mode texturefeatures for ROI sizes of 4 mm2 and 9 mm2, for comparison with ROIs of 1 mm2.162.2.3 Classification and Estimation of Cancer LikelihoodSupport vector machine (SVM): SVM is a widely used maximum margin classifier.The soft margin SVM classifier is a hyperplane of the form wφ(xi)+b, where φ isthe function that maps the feature vector, xi, to a higher dimensional space and wand b are determined to minimize [75]:12×wT w+CN∑i=1ξi (2.1)subject toyi(wTφ(xi)+b)≥ 1−ξi (2.2)where C > 0 is the regularization or penalty parameter that minimizes the error bycontrolling the trade-off between the slack variable penalty and the margin, andξi ≥ 0 are the slack variables that provide flexibility when fitting the data by per-mitting incorrect classification of noisy and difficult data points [75]. The functionφ(.) maps the data to a higher dimensional space. This new space is defined byits kernel function. It can be shown that in the SVM formulation, one does notneed an explicit expression for the kernel function and the SVM optimization anddecision hyperplane are defined fully given the dot product format of the kernelK(xi,x j) = φ(xi)T φ(x j).The classification problem can also be written in the form of its dual repre-sentation in terms of the kernel function. In this representation, the class label isdetermined based on the sign of y(x) in the following equation:y(x) =N∑i=1aitiK(x,xi)+b (2.3)where y(x) is the model prediction of input x, ti is the true label for the supportvector xi, ai is the Lagrange multiplier used to convert the maximum margin SVMoptimization problem from 2.1 to its dual representation. The solution to the SVMtraining is a quadratic programming problem that solves for ai’s.The soft margin SVM algorithm implemented in the scikit-learn Python pack-age and reported in [76] was used for tissue classification. The radial basis func-tion (RBF) kernel was employed. The values of the RBF exponent (γ), which de-17termines the range of support, and the parameter C, which controls the trade-offbetween the slack penalty and the margin in the soft margin SVM, were chosenbased on a grid search through leave-one-subject-out cross-validation.Cancer likelihood is estimated using Platt’s algorithm [77] as follows. Assumethat the SVM hyperplane obtained after training is wφ(x)+b where x is the featurevector and φ is the function that maps the feature vector to a higher dimensionalspace. Cancer likelihood (Lc) is computed by mapping the distance of each testsample to the decision boundary using a sigmoid function of form:Lc = (c|(wφ(xi)+b)) =11+ exp(α(wφ(xi)+b)+β )(2.4)Where c stands for cancer (class). Maximum likelihood estimation from the obser-vations for which the true labels are known (training data) was used to calculate thevalues of the parameters α and β . See [75] for details and [78] for implementation.Random Forests: Random Forests is a robust classification algorithm employedfor medical image processing applications [79, 80]. Random Forests algorithm is acollection of decision trees that exploits the bias variance trade-off. The underlyingconcept behind the algorithm is that by taking the prediction of each decision treeand averaging it across the ensemble/forest the high variance in the prediction ofdecision trees can be reduced. The Random Forests algorithm can be described as:1. For 1 to T trees:(a) Draw a bootstrap (bagging) sample from the training data(b) Recursively repeat the steps below at each terminal leaf, until the de-sired depth is reached, to construct a random tree from the bootstrappeddata.i. Randomly select a subset of n features from f featuresii. Select the split point that results in maximum information gainalong each n featureiii. Split the leaf into two child leaves2. Output the forest of T trees18The objective function at each leaf of the tree is I = I(S j,θ), where S j is the jthnode, θ is the leaf parameter or the split point, and I is the information gain. Theinformation gain is calculated asI(S,θ) = H(S)− ∑iε{L,R}|Si||S|H(Si) (2.5)where, H(S) is Shannon’s entropy calculated asH(S) =−∑cεCp(c)log(p(c)) (2.6)In the above equation C is the output/classes. From the n randomly selected fea-tures, the feature that results in the highest information gain is used at the root nodeand the best split point is also calculated with the objective of maximizing the infor-mation gain. During testing, the sample is pushed down each tree simultaneouslyuntil it reaches the corresponding leaf.Random Forests implementation within the scikit-learn Python package wasused [76]. Forest class a posteriori probabilities or malignancy probabilities p(c|v),is calculated as the average of all tree a posteriori probabilities pt(c|v) and can bedescribed as:p(c|v) =1TT∑tpt(c|v) (2.7)where v is the input data, c stands for cancer (class), and T is the number of decisiontrees in the forest. The a posteriori class probability pt(c|v) is calculated at eachleaf of the decision tree as the fraction of the number of samples with the majoritylabel over the total number of samples. The label assigned to the test data is basedon the forest class a posteriori probability [81]. The algorithm parameters wereoptimized by performing a grid search and the decision trees were constructedshallow to avoid over-fitting. The classification algorithm implemented in scikit-learn package closely follows the method described in [79] and [80].Feature and Model Selection: To use the most informative features and reducethe dimensionality of the data, only a subset of the features were chosen for clas-sification. It is important to note that the Random Forests algorithm provides abuilt-in method for ranking features in terms of their ability to separate samples19from the two classes. This is not the case for SVM. Therefore, our approach tofeature selection was different with the two classifier models.In case of the SVM classifier, we used recursive feature elimination (RFE).The RFE algorithm, originally reported in [82], recursively eliminates the featureswith the smallest contribution to an estimated linear maximum margin model andis shown to outperform correlation-based methods in gene selection [82]. In thismethod, features are eliminated recursively based on their corresponding weight ina linear SVM classifier of the form y = wT x+b. Initially, the model is trained onall the features and weights are calculated for each one of them. Then the featurewith the smallest absolute weight value is eliminated. This process is then repeatedrecursively until the desired number of features is reached [82].In Random Forests algorithm, feature importance is estimated as follows: eachdecision tree is created using a bootstrap sample of the original data. Hence, theleft out data is labelled as out-of-bag (OOB) samples and can be used as “test data”to estimate feature importance. In this step, first all OOB samples are propagatedthrough the trained model and the accuracy is recorded. Then the values of onefeature for OOB samples are randomly permuted and the OOB samples with thispermuted feature value are propagated down the trees. The expectation is that thepermutation results in decreased accuracy. The permutation of the “important”features will cause a larger reduction in classification accuracy. The differencebetween average ensemble accuracy with and without permutation provides a rawmeasure of feature importance. The features were ranked based on this importancemeasure. The choice of the number of features to select for classification wasconsidered as one of the model parameters. All model parameters, which includedthe depth of trees, the number of trees, and the number of features were optimizedthrough a grid search. Leave-one-subject-out cross-validation was used for thisgrid search.2.2.4 Evaluation of ClassificationThe classification label was compared with the biopsy result, which was used as theground truth, to compute the classification accuracy. We assess the performance ofthe classifiers by calculating the AUC. The ROC curves were generated using the20method described in [83] and the AUC was estimated using the trapezoidal rule.We also calculated the standard errors and p-values for the AUCs using the Hanleyand McNeil method reported in [84].In order to evaluate the performance of other groups of features as opposedto RF time series, we trained an SVM classifier using parameter selection thatincluded only texture and attenuation parameters. All other aspects of the methodwere similar to the time series experiment.2.3 Results2.3.1 Feature SelectionThe feature selection algorithm showed that using three features resulted in thelowest cross-validation error in both SVM and Random Forests. Features 2, 3, and7 were the best performing for classification using SVM and features 2, 5, and7 yielded the highest importance score for classification using Random Forests.It should be noted, that the fractal dimension feature extracted from the RF timeseries, feature 7, was the best overall. Note that for both classifiers the final set offeatures only included RF time series features.2.3.2 Classification ResultsTable 2.3 shows the BI-RADS score, tissue type, classification result, and classifi-cation accuracies for all the 22 subjects (IDC stands for invasive ductal carcinoma,FA stands for Fibroadenoma, Infl. stands for Chronic Inflammation, and NS standsfor Nodular Sclerosis). The tissue type is not available for subject 12 because thecytology result was unknown due to low cellularity.Both of the classification algorithms, Random Forests and SVM, correctly clas-sified 18 out of 22 subjects. We define success as correct labelling of at least 60%of the 1 mm2 ROIs from the lesion. Cancer likelihood for each ROI, which is anumber between 0 and 1, was translated to a colormap applied to the lesion areaon the B-mode image. The result for several cases is illustrated in Figure 2.9 andshows the utility of the method in visualizing the likelihood of malignancy. As canbe seen in Table 2.3, Subjects 3, 7, 10, and 16 were incorrectly classified. Further2100.20.40.60.81Malignant BenignFeature0100.20.40.60.81Malignant BenignFeature0200.20.40.60.81Malignant BenignFeature0300.20.40.60.81Malignant BenignFeature0400.20.40.60.81Malignant BenignFeature0500.20.40.60.81Malignant BenignFeature0600.20.40.60.81Malignant BenignFeature0700.20.40.60.81Malignant BenignFeature0800.20.40.60.81Malignant BenignFeature09Figure 2.5: Box plots of the nine RF time series features. Table 2.1 providesa description of the features. From the box plots, it is apparent that thebest features selected by the RFE algorithm show a difference in theclass distribution. As can be seen in the figure, ”Feature07” is the bestfeature that differentiates the malignant samples from the benign, and itwas also selected as the best feature by our feature selection method.investigation of these subjects showed calcification inside the lesion for Subject3 and poor lesion boundaries for Subjects 7 and 16. It should also be noted thatSubject 7, one of the incorrectly classified malignant cases, was the smallest lesionin our dataset. Figure 2.10 shows the malignancy maps for these subjects.22164 Texture Features164 Texture Features 0 20 40 60 80 100 120 140 160020406080100120140160−1−0.8−0.6−0.4−0.200.20.40.60.81Figure 2.6: Correlation plot of the 164 texture features extracted from the B-mode image. From the figure, it is observable that most of the featuresare highly correlated. High positive correlation suggests the need forfeature selection prior to classification.As shown in Figure 2.7, the AUC using SVM was 85.6% (95% confidence in-terval (CI): 83.9%–89.8%) and using Random Forests was 81.3% (95% CI: 78%–85%). We also calculated the classification accuracies for each test subject andthese can be found in Table 2.3. The mean classification accuracy for each sub-ject was calculated as 76% for SVM and 74% for Random Forests. As depictedin Table 2.2, using a cut-off threshold of T hc=0.5, classification using SVM showssensitivity of 82% and specificity of 80%. However, a drop in sensitivity (73%) anda rise in specificity (88%) was seen when using Random Forests. Our dataset con-sists of mainly benign lesions and hence the majority of the training data belongsto the benign class. The decrease in sensitivity could be due to this class imbal-ance problem [85, 86]. Another possible cause of the lower classification accuracy23Figure 2.7: ROC curve for SVM and Random Forests using three features.using the Random Forests algorithm is the small number of selected features. Thepower of Random Forests is in random sampling of features, which is most usefulin high-dimensional feature vectors [80].Using the statistical tests proposed in [84], the standard errors were calculatedas 0.016 and 0.018 for AUC using SVM and Random Forests, respectively. Us-ing the same method, and performing the z test, the p-value was calculated to be0.008, hence, the difference in performance of the two classifiers was statisticallysignificant (p < 0.05) [84].24Figure 2.8: Classification results for the 22 subjects. As can be seen in thefigure, five out of seven malignant subjects were correctly classified and13 out of 15 benign subjects were correctly classified.Figure 2.9: The malignancy map created by plotting the value cancer likeli-hood (Lc) of the ROIs overlaid on the B-mode image. The images on thetop are from malignant subjects, and the ones on the bottom are benignsubjects.25Table 2.2: Sensitivity, specificity at a cut-off T hc=0.5, and area under ROCcurve.Algorithm Sensitivity Specificity AUC (95% CI)SVM 82% 80% 85.6% (83.9%–89.8%)Random Forests 73% 88% 81.3% (78.0%–85.0%)Table 2.3: Classification results for the 22 subjects.Subject BI-RADS Tissue Classification Classification AccuracyType Result SVM Random Forests1 5 IDC Malignant 100% 96%2 5 IDC Malignant 94% 93%3 5 IDC Benign 0% 0%4 4B IDC Malignant 96% 71%5 4C IDC Malignant 100% 100%6 4C IDC Malignant 100% 100%7 4B IDC Benign 0% 0%8 4A FA Benign 100% 100%9 3 FA Benign 86% 73%10 4 FA Malignant 0% 0%11 4B FA Benign 100% 100%12 4A Unknown Benign 88% 83%13 4B FA Benign 96% 93%14 4A FA Benign 100% 100%15 4B Infl. Benign 100% 100%16 4A FA Malignant 0% 0%17 4A FA Benign 69% 94%18 4A FA Benign 100% 100%19 4A Cyst Benign 60% 53%20 4A FA Benign 75% 73%21 4A FA Benign 100% 100%22 4A NS Benign 100% 100%The effectiveness of the RF time series method is evident from the fact that onlyRF time series features were selected by our systematic ranking method. Classi-fication using the best three single-frame RF spectral and B-mode texture featureswas also performed for comparison with time series features. The three selected26Figure 2.10: The malignancy maps for the subjects that the classifiers did notcorrectly label. On the top are two incorrectly classified malignantsubjects (Subjects 3 and 7), and the ones on the bottom row are theincorrectly classified benign subjects (Subjects 10 and 16).features in this experiment were the intercept and slope from the attenuation group(Features 8 and 9) and mean intensity from the texture group (Feature 10). Thiscombination resulted in an AUC of 68%. However, these parameters provide abenefit. Specifically, the classifier trained on texture and attenuation parameterswas successful in classifying both of the benign subjects that were mis-classifiedby the time series method. Using the best three texture and attenuation features theclassification accuracies for subjects 10 and 16 were calculated as 70% and 91%.This was, however, at the cost of reduced sensitivity and overall accuracy.27Figure 2.11: Scatter plot showing the ROI samples from the 22 subjects usingfeatures 2 and 3. Red and blue markers represent malignant and benignROI samples from 22 subjects.We suspected that the poor classification accuracy of the single-frame RF spec-tral features and the B-mode texture features could be due to the small (1 mm2) ROIsize, hence, we increased the ROI size to 4 and 9 mm2 and calculated the AUCs as69.4% and 70.2%, respectively. The difference in AUCs based on ROI sizes wasnot significant because the p-value was calculated as 0.91. We did not go above9 mm2 for the ROI size because that would have resulted in loss of data from twosubjects due to small lesion size.28Figure 2.12: SVM decision function plot for the malignant subjects. The redregion is the decision boundary for the malignant ROI samples and thecyan region is the decision boundary for the benign ROI samples. Ascan be seen in the figure, malignant subjects, 3 and 7, lie in the benignregion and are therefore misclassified by the classification algorithms.See Figure 2.11 for distribution of the ROI samples.2.4 Discussion and ConclusionThere exists a need to reduce the number of false positives based on a BI-RADSscore, hence, augmentation of the current breast cancer diagnosis methods is apressing need. In this work we presented an ultrasound-based method for classi-fication of malignant breast lesions. This method uses RF time series parameterswithin a machine learning framework. A set of 21 parameters extracted from RFtime series analysis, B-mode texture, and single-frame RF spectral analysis wereranked for their importance with two different feature ranking approaches. In bothfeature-ranking algorithms, all the three best performing features were from the RF29Figure 2.13: Random Forests decision function plot for the malignant sub-jects. The red region is the decision boundary for the malignant ROIsamples and the cyan region is the decision boundary for the benignROI samples. As can be seen in the figure, malignant subjects, 3 and 7,lie in the benign region and are therefore misclassified by the classifi-cation algorithms. See Figure 2.11 for distribution of the ROI samples.time series group. Using these three features, we obtained an AUC of 85.6% and81.3% on SVM and Random Forests classifiers, respectively. The best performingsubset of features that were not from RF time series resulted in an AUC of 68%.The performance of B-mode texture features in breast cancer classification waspoor in our study. However, it should be noted that the B-mode images used inthis work were reconstructed from the RF signals offline. The B-mode imagesfrom the commercial ultrasound machines are filtered and optimized in terms ofdynamic range and have a higher quality compared to the B-mode images usedhere. Therefore, the reported performance of the B-mode texture features could be30Figure 2.14: SVM decision function plot for the benign subjects. The redregion is the decision boundary for the malignant ROI samples andthe cyan region is the decision boundary for the benign ROI samples.As can be seen in the figure, benign subjects, 10 and 16, lie in themalignant region and are therefore misclassified by the classificationalgorithm. Subject 10 is an outlier in the dataset and significantly ef-fects the classification results. See Figure 2.11 for distribution of theROI samples.31Figure 2.15: Random Forests decision function plot for the benign subjects.The red region is the decision boundary for the malignant ROI samplesand the cyan region is the decision boundary for the benign ROI sam-ples. As can be seen in the figure, benign subjects, 10 and 16, lie inthe malignant region and are therefore misclassified by the classifica-tion algorithm. Subject 10 is an outlier in the dataset and significantlyeffects the classification results. See Figure 2.11 for distribution of theROI samples.32potentially improved by using the B-mode images produced by the scanner.RF time series method is susceptible to patient motion. The time series used inthis work were obtained over less than 2 seconds. We found that by asking the sub-jects to hold their breath, the amount of motion could be minimized. An ultrasoundtransducer stabilization mechanism could also reduce the operator-introduced mo-tion and increase the length of motion-free time series, which is linked to improvedtissue classification in previous work [37]. Our measurement based on [68] showedthat the amount of displacement in the RF data had the highest value in the misclas-sified malignant subjects. Also, our dataset includes mostly benign subjects whichintroduces a class imbalance problem. This affects the classification performance,specially the sensitivity. We need a larger dataset with improved representation ofcalcification and malignancy. As the size of the dataset increases, we anticipateimproved performance.The physical basis of the tissue typing capabilities of the RF time series methodhas been the topic of a few previous works. In [87], a model was developed torelate the variations of the ultrasound backscattering to the variations in tissue tem-perature and speed of sound that take place during the RF time series scanningprocedures. The measurements of the variations in the ultrasound backscatteringwere then used in a tissue classifier. Other reports show that an increase in theframe rate and the acoustic power of the ultrasound beam result in an improvedaccuracy of tissue classification [37]. To prove the tissue typing capabilities of ul-trasound RF time series we have conducted experiments on animal tissue to studythe effects of ultrasound imaging parameters on tissue classification. The results ofthese experiments can be found in Chapter 4.Accurate malignancy maps could eventually be used as a diagnosis methodwhich would significantly reduce the negative biopsy outcome. Our proposedmethod can be part of the overall solution for multiparametric ultrasound analy-sis of breast cancer. We argue that RF time series can be a practical componentof that approach without the need for additional equipment. This method is alsoindependent of the contour shape and the potentially subjective process of segmen-tation.332.5 Chapter SummaryThis chapter reported the use of ultrasound RF time series analysis as a methodfor ultrasound-based classification of malignant breast lesions. The RF time seriesmethod is versatile and requires only a few seconds of raw ultrasound data withno need for additional instrumentation. Using the RF time series features, and amachine learning framework, cancer maps were generated, from the estimated can-cer likelihood, for decision support in biopsy recommendation. These maps depictthe likelihood of malignancy for regions of size 1 mm2 within the suspicious le-sions. We reported an area under receiver operating characteristics curve of 85.6%using support vector machines and 81.3% using Random Forests classification al-gorithms, on 22 subjects with leave-one-subject-out cross-validation. Changing theclassification method had an insignificant effect on the accuracy which indicates therobustness of the tissue typing method. The findings of this chapter suggest thatultrasound RF time series, along with the developed machine learning framework,can help in differentiating malignant from benign breast lesions, subsequently re-ducing the number of unnecessary biopsies after mammography screening.34Chapter 3Ultrasound-based Prediction ofProstate Cancer in MRI-guidedBiopsy3.1 IntroductionProstate cancer is the most common type of solid tumor, and the second lead-ing cause of cancer-related deaths in North American and European men. Earlystage prostate cancer, which represents the majority of cases diagnosed today, hasmany therapy options, including surgery, radiation therapy, brachytherapy, thermalablation, and active surveillance. Selection of the optimal therapy and therapeu-tic dosage are chiefly determined by diagnosis and staging. Definitive diagnosisof prostate cancer requires core needle biopsy, typically guided by TRUS. Cur-rent biopsy regimens involve systematic sampling of the prostate from eight ormore predefined anatomical locations, followed by histopathological evaluation ofthese samples. The biopsy regimen is scaled to the prostate gland based on itssize and using nomograms but otherwise not tailored to the individual. TRUS-guided biopsy has rather poor sensitivity, with positive predictive values between40-60% [88]. Improved cancer yield can be achieved if patient-specific targetingis combined with systematic sampling. However, this is not feasible using TRUS35alone.In order to enable patient-specific targeting, other modes of ultrasound imag-ing [36], such as RF data analysis [27], elastography [28], and Doppler imag-ing [30] have been explored. These technologies, individually, have not entirelyproved successful in accurate identification of high grade cancer.MRI has been used as an alternative modality to improve high grade prostatecancer yield [18]. Guidelines for structured reporting of prostate cancer assess-ments based on multi-parametric MRI have been developed, involving simulta-neous examination of T2-weighted, DCE T1-weighted, and DWI sequences [23].MRI-guided biopsy is, however, difficult, costly, time-consuming and not avail-able in many parts of the world [22]. Fusion of ultrasound and MRI has beenused to improve prostate cancer detection by enabling targeting of the cancer focipre-determined in MRI during TRUS-guided biopsy [18]. Biopsy core locationsdetermined in MRI are translated to patient coordinates using pre-procedure 3DTRUS and its registration to MRI [24]. 3D TRUS to MRI registration requireseither sophisticated mechanical systems [25] to guide the biopsy needles or, if per-formed by software only [26], does not fully account for patient motion or organdeformation occurring during biopsy.In Chapter 1 we introduced ultrasound RF time series as a tissue classificationmethod. An RF time series is formed when ultrasound RF frames are capturedin time from a stationary tissue location. In this thesis, we propose to use ultra-sound RF time series to complement MR-targeted biopsy procedures by providingcancer likelihood maps around MRI targets during biopsy. We envision that thissolution should increase positive cancer yield in both MR-targeted and/or TRUS-guided biopsy procedures. It will also provide an opportunity to correct for mis-registrations of MR and TRUS images prior to sampling the tissue.In the proposed solution, RF time series features have been used within aninnovative computational framework that combines unsupervised clustering of thedata with supervised classification. We use the histopathology of the biopsy coresfor evaluation of cancer detection. Cancer likelihood maps are also shown thathighlight the distribution and the likelihood of cancerous tissue within the biopsycores. In a single centre feasibility trial with data obtained from 14 subjects at 18biopsy targets, we are able to predict the pathology of MRI-identified targets with36high specificity and sensitivity.3.2 Materials and Methods3.2.1 Data AcquisitionUltrasound RF time-series data is acquired on a Philips iU22 US scanner duringMRI-guided targeted TRUS biopsies performed at the National Institutes of HealthClinical Center (NIH-CC, Bethesda, MD) using the Philips UroNav platform. Fortargeted biopsy, pre-acquired T2-weighted MRI images are automatically fusedwith real-time TRUS images of the prostate [22]. Initially, the desired targets aredelineated on the T2-weighted MRI image by a clinician based on the examinationof four multi-parametric MR images: T2-weighted, DWI, DCE, and MR spec-troscopy. At the beginning of the biopsy procedure, a series of electromagneticallytracked two-dimensional (2D) TRUS images of the prostate are acquired from baseto apex. Next, a 3D US volume is reconstructed based on electromagnetic (EM)tracking data and registered to the MRI scan in the UroNav software. Followingthe registration of US and MR volumes, the targeted locations for biopsy are trans-formed to the EM coordinate frame. During the biopsy, the clinician navigatesthrough the prostate volume to reach the desired target location for acquiring acore. Immediately prior to taking the biopsy, the clinician holds the TRUS trans-ducer steady for 4-5 sec to acquire RF time series data. Typically, 100 frames ofRF time series data are acquired from each biopsy core. RF data is obtained priorto one, and in some cases, two biopsies of the MR-identified targets.Ultrasound RF time series data is used from 18 biopsy cores of 14 subjects.Although RF time series data is collected in the axial plane, two biopsies are takenfrom axial and sagittal planes for each subject from the same location. The record-ing of the RF data and acquisition of the biopsy core are performed in sequence, notsimultaneously, to avoid the appearance of the needle in the images. As a result,hand motion maybe present in some cases, between data and biopsy acquisitionas well as during RF data recording. A quality control step is necessary to ob-tain a dataset with reliable reference label. In this step, we only choose to includesubjects for which the histopathology of the axial and sagittal biopsies agree, and37Figure 3.1: Philips UroNav platform (left) and a Philips iU22 ultrasound ma-chine (right).no excessive motion is present during RF time series acquisition. In our data, 10biopsy cores are cancerous with Gleason scores above 6 and tumor areas >40%.Eight biopsy cores are benign with consistent histopathological information.3.2.2 Feature ExtractionRegions of Interest (ROIs): For each registered biopsy target, we analyze an area of2 mm×10 mm in the lateral and axial directions, respectively, along the projectedneedle path in the RF data, and centred on the target. The width of this area isclose to the width of the biopsy core. The length of the biopsy core is typicallylarger than the 10 mm considered here; however, to account for mis-registrationerrors and possible hand-motions, we use a conservative estimate in this study.The selected 2×10 mm area is divided into 20 ROIs of size 1×1 mm resulting atotal of 360 ROIs from all biopsy cores. For each ROI, we calculate the featuresdescribed below.Features: Nine tissue typing parameters are extracted using the spectral, fractal,38Figure 3.2: B-mode image showing the ROI selection. ROIs were selectedalong the biopsy needle trajectory as shown in the figure.and wavelet analysis of the RF time series data. Each RF time series contains 96sequentially acquired frames of each RF sample of the imaging plane. We computethe spectrum of the zero-mean, hamming windowed, time series of an RF sampleand average the values over an ROI. Summation of the spectrum in four equally-spaced frequency bands constitute features 1-4 [32]. The intercept and slope of thefitted line to the spectrum in the entire frequency range are features 5 and 6. Fractaldimension of the time series is computed using Higuchi’s method and averagedover an ROI as feature 7 [32]. We also calculate the central frequency (CF) ofthe spectrum as the mean of the spectrum bandwidth of the time series of an RFsample. The mean of the CF values (MCF) over an ROI is used as feature 8 [40].Finally, we apply the discrete wavelet transform to the ultrasound RF time seriesof each RF sample using Daubechies 4 filter bank, where the signal is decomposedinto approximation and detail coefficients at each decomposition level. The firstapproximation coefficient is computed for each RF sample in the imaging plane atthe coarsest level (n=3) of decomposition and averaged over each ROI as feature9 [40].393.2.3 The Proposed Classification FrameworkFeature selection: Feature selection is performed using Recursive Feature Elimi-nation (RFE), prior to classification to identify the optimal combination of the ninefeatures described above for cancer detection. In this method, features are elimi-nated recursively based on their corresponding weight in a linear SVM classifier.Initially, the model is trained on all the features and their weights are calculated.Then, the feature with the smallest absolute weight value is eliminated. This pro-cess is repeated recursively; the number and combination of features resulting inthe highest classification accuracy are used as the stopping criteria. In our case, thecombination of two features resulted in the highest classification accuracy.Calculate  RF  ⬬e  series  features  for  each  ROI  K-­‐means  Clustering  of  ROIs  Train  a  classifier  model  on  each  cluster  Test  using  the  appropriate  cluster-­‐classifier  model.  Figure 3.3: An overview of the classification framework employed in thiswork. Clustering of the data is performed prior to classification to elim-inate the affect of outliers and improve the classification accuracy.Classification: Even though our biopsy cores are assigned to cancer or be-nign pathologies, the selected tissue types are heterogeneous within these classesand could potentially be differentiated based on other structural differences. Oneapproach to overcome “within class” differences is to first cluster the ROIs in anunsupervised manner. This could result in identifying the outliers of each classfrom the main distribution of the class. A cluster-specific classifier can then beused to differentiate cancerous and benign tissue in a supervised manner.Experiments: We follow a leave-one-subject-out cross-validation strategy. Here,we train a classifier using the features extracted from the cancerous and benign40ROIs of biopsy cores from 13 subjects and test on the features extracted from theROIs of an unseen subject. In the first step of the process, ROIs from all 13 train-ing subjects are clustered into two groups using k-means algorithm. Within eachcluster, we train a Support Vector Machine (SVM) classifier to separate cancerousfrom benign ROIs. The next step constitutes testing, where we first assign the ROIsof the unseen subject to one of the clusters based on their Euclidean distances fromthe centroids of the clusters. The ROIs of the test subject are then classified usingthe classifier corresponding to their respective clusters. This process is repeated 14times where every subject is left out for testing once. If in any of these leave-one-subject-out trials, a resulting cluster after the k-means step is over 90% imbalanced(over 90% benign or cancer), we do not train a classifier for that cluster and thelabel of test samples are determined based on majority voting in that cluster. Inorder to ensure that our process is not tailored to one type of classifier, we alsouse a Random Forests classifier and report our results using the two classificationmethods.K-means clustering, SVM and Random Forests algorithms are implementedin the Scikit-learn machine learning package [76]. In addition to the binary classlabels, we also estimated the cancer likelihood [77] for the biopsy core of each sub-ject. The hyper-parameters that need to be determined for the classifiers includedthe Radial Basis Function (RBF) exponent and the soft margin penalty coefficientfor SVM, and the number and depth of the trees in the Random Forests. These aretuned using a grid search approach.3.3 ResultsThe RFE feature selection process was repeated for every leave-one-subject-outexperiment. It consistently isolated features 3 and 4 as the combination of featuresthat result in the highest classification accuracy between cancerous versus benigntissue. These are both spectral parameters of the RF time series. Henceforth, weonly use these two features in clustering and classification of the biopsy cores.Figure 3.7 shows the two clusters that are created by k-means for ROIs from allsubjects. 190 out of 200 malignant ROIs are assigned to cluster 1 and 125 out of160 benign ROIs are assigned to cluster 2. In other words, 95% of all cancerous4100.20.40.60.81Malignant BenignTime Series 100.20.40.60.81Malignant BenignTime Series 200.20.40.60.81Malignant BenignTime Series 300.20.40.60.81Malignant BenignTime Series 400.20.40.60.81Malignant BenignTime Series Intercept00.20.40.60.81Malignant BenignTime Series Slope00.20.40.60.81Malignant BenignFractal Dimension00.20.40.60.81Malignant BenignMCF00.20.40.60.81Malignant BenignWaveletFigure 3.4: Boxplots of features.samples and 78% of all benign samples are grouped in clusters 1 and 2, respec-tively. Based on this observation, and in order to maximize the number of trainingdata per cluster, we limit the number of clusters to two.The ROC curves are found in Figure 3.8. The area under the curve is 91.5% and87.4% (95% CI: 85.7%–92.7%) for SVM and Random Forests methods, respec-tively. The difference in the area under the curves, of SVM and Random Forests,was statistically significant (p < 0.05). Colormaps that depict the cancer likelihoodof ROIs in each of the 18 biopsy cores are illustrated in Figure 3.10. The likelihood42 Feature01 Feature02 Feature03 Feature04 Feature05 Feature06 Feature07 Feature08 Feature09Feature01Feature02Feature03Feature04Feature05Feature06Feature07Feature08Feature09−1−0.8−0.6−0.4−0.200.20.40.60.81Figure 3.5: Correlation plots of the RF time series features. As can be seenin the figure, all the time series features are positively correlated, exceptfor features 6 and 9. Due to high correlation among the features, weperform feature selection before classification.threshold to label an ROI cancerous in the cancer likelihood maps is chosen to be60%. It is noteworthy that if we eliminate the clustering step and perform classi-fication with all training samples, we obtain an area under the curve of 86.9% and87.1% for SVM and Random Forests methods, respectively.Table 3.1 shows the percentage of the number of ROIs predicted as cancerousin each core, found in test samples in the leave-one-subject-out classification. Thetwo different columns report the outcome for our method using SVM and RandomForests as classifiers. Using the SVM classifier, the percentage of cancer foundin all benign cores is 45% or smaller and in five out of eight benign subjects thisnumber is zero. In the positive biopsy cores, we notice that the predicted percentageof cancer is above 60% using the SVM classifier.43Figure 3.6: Histograms of the features. The best selected features, features 3and 4, are highlited in red.3.4 Discussion and ConclusionWe present a machine learning framework, consisting of supervised and unsuper-vised learning approaches, that uses RF time series analysis for the prediction of thehistopathology of MR-guided targeted prostate biopsies. In a leave-one-subject-outstudy with data obtained from 18 biopsy cores in 14 subjects, we are able to ac-curately predict the pathology of MRI-identified targets with high specificity andsensitivity. In ROIs as small as 1 mm×1 mm, and using only two spectral fea-tures of RF time series, an area under ROC curve of 91.5% is achieved. Usingk-means clustering, we show that these two features are able to separate cancer-ous and benign biopsy cores. Following classification, we calculate similar areaunder the ROC curve independently with SVM and Random Forests; this points tothe stability of the proposed framework for tissue classification. We also present44Figure 3.7: Clustering performed on all 360 training samples.colormaps that depict the cancer likelihood of ROIs in biopsy cores. These mapsclosely match the histopathology results of each biopsy core. As Table 3.1 shows,we report low cancer probabilities for all benign cores; specifically we predict zerocancer likelihood for five out of eight benign cores. In other words, 63% of thenegative biopsies could have been avoided had we known the cancer likelihood ofthat area using RF time series during biopsy. In terms of sensitivity, as is observedin Table 3.1, we report at least 60% (mainly 70% and up) cancer likelihood for allpositive cores.Our results demonstrate that RF time series can be used to complement MR-targeted biopsy procedures, by providing cancer likelihood maps around MRI tar-gets during biopsy. Our proposed method could potentially increase positive canceryield in both MR-targeted and/or TRUS-guided biopsy procedures. It could alsobe used to compensate for mis-registrations of MR and TRUS images prior to sam-pling the tissue for MRI guided prostate biopsies.45Figure 3.8: ROC curves for SVM and Random Forests (each performed afterclustering).A limitation of our study is the size of the dataset. This is partly due to ourconservative quality control step where we drop data from targets with conflictingpathology results in axial and sagittal planes. In addition, to minimize the impactof registration and targeting error on our analysis, we only choose ROIs in 10 mmlength of the RF data centred around the target along the needle trajectory. A typ-ical biopsy core could be as long as 18 mm. Data acquisition for a large clinicalstudy is ongoing; the aim is to also incorporate a detailed histopathology report-46Figure 3.9: Classification results for the 18 bipsy cores. As can be seen in thefigure, all biopsy cores were correctly classified.Figure 3.10: Cancer likelihood colormaps of the 18 biopsy cores from 14subjects with leave-one-subject-out cross validation using the best twoRF time series features. Clustering and SVM classification is used.47Table 3.1: SVM and Random Forests cancer probabilities.Subject Biopsy Biopsy Gleaso Percentage of CancerCore Result Score SVM Random Forests1 Core 1 Adenocarcinoma 7 100% 75%2 Core 2 Adenocarcinoma 8 90% 85%Core 3 Adenocarcinoma 8 70% 70%3 Core 4 Adenocarcinoma 6 85% 70%4 Core 5 Adenocarcinoma 9 95% 95%Core 6 Benign 0 45% 25%5 Core 7 Adenocarcinoma 8 75% 90%Core 8 Adenocarcinoma 8 95% 100%6 Core 9 Adenocarcinoma 7 75% 70%7 Core 10 Adenocarcinoma 7 100% 100%8 Core 11 Adenocarcinoma 7 60% 70%9 Core 12 Benign 0 0% 60%10 Core 13 Benign 0 0% 0%11 Core 14 Benign 0 0% 0%Core 15 Benign 0 30% 10%12 Core 16 Benign 0 0% 0%13 Core 17 Benign 0 15% 10%14 Core 18 Benign 0 0% 0%ing scheme where the direction of the cancer in a core is marked and results arereported in quarters along the biopsy core. We expect a larger dataset and moreaccurate mapping of histopathology to RF time series would further improve theresults.3.5 Chapter SummaryIn this chapter, an in vivo clinical feasibility study for ultrasound-based detectionof prostate cancer in MRI selected biopsy targets was reported. Spectral analysisof a temporal sequence of ultrasound RF data reflected from a fixed location inthe tissue results in features that can be used for separating cancerous from benignbiopsies. Data from 18 biopsy cores and their respective histopathology were usedin an innovative computational framework, consisting of unsupervised and super-48Figure 3.11: Scatter plot showing the ROI samples from the 18 biopsy coresusing features 3 and 4. Red and blue markers represent malignant andbenign ROI samples from 18 cores.vised learning, to identify and verify cancer in regions as small as 1 mm×1 mm. Inleave-one-subject-out cross validation experiments, an area under ROC of 91.5%(95% CI: 90.6%–95.7%) was obtained for cancer detection in the biopsy cores.Cancer likelihood maps that highlight the predicted distribution of cancer alongthe biopsy core, also closely match histopathology. The results of this chapterdemonstrate the potential of the RF time series to assist patient-specific targetingduring prostate biopsy.49Figure 3.12: SVM decision function plots for the malignant cores. The de-cision plots shown here are without clustering. The red region is thedecision boundary for the malignant ROI samples and the cyan regionis the decision boundary for the benign ROI samples. It is apparentin the above figure, some ROI samples from cores, 3, 8, 9, and 11 arein the benign region and without clustering would have been misclas-sified. These decision functions (plotted without clustering) show theutility of clustering prior to classification. See Figure 3.11 for distri-bution of the ROI samples.50Figure 3.13: Random Forests decision function plots for the malignant cores.The decision plots shown here are without clustering. The red regionis the decision boundary for the malignant ROI samples and the cyanregion is the decision boundary for the benign ROI samples. It is ap-parent in the above figure, some ROI samples from cores, 3, 8, 9, and11 are in the benign region and without clustering would have beenmisclassified. These decision functions (plotted without clustering)show the utility of clustering prior to classification. See Figure 3.11for distribution of the ROI samples.51Figure 3.14: SVM decision function plots for the benign cores. The decisionplots shown here are without clustering. The red region is the decisionboundary for the malignant ROI samples and the cyan region is thedecision boundary for the benign ROI samples. As can be seen in theabove figure, most of the benign ROI samples lie in the benign region(cyan). Some ROI samples from core 6 are in the malignant regionand hence the classification accuracy for core 6 is lower than the otherbenign cores. See Figure 3.11 for distribution of the ROI samples.52Figure 3.15: Random Forests decision function plots for the benign cores.The decision plots shown here are without clustering. The red regionis the decision boundary for the malignant ROI samples and the cyanregion is the decision boundary for the benign ROI samples. As canbe seen in the above figure, most of the benign ROI samples lie inthe benign region (cyan). Some ROI samples from core 6 are in themalignant region and hence the classification accuracy for core 6 islower than the other benign cores. See Figure 3.11 for distribution ofthe ROI samples.53Chapter 4Experiments to Provide Evidenceon the Physical Basis of TissueClassification using RF TimeSeries4.1 IntroductionUltrasound RF time series method is a data-driven approach that has yielded con-sistent results in classification of animal tissue [37], prostate cancer ex-vivo [32],prostate cancer in-vivo [39, 40], breast lesions [66], and in monitoring of tissueablation [41]. While we have not discussed the potential reasons for improvedperformance obtained from the use of a sequence of frames, several hypothesesdescribing the physical phenomenon exist and have been explored elsewhere.It has been shown experimentally and in simulations, that the temperature in-crease in controlled irradiation of tissue with RF time series can partly explain thephenomenon [87]. Also, increased energy delivered to the tissue through increasedframe rate or transmit power result in improved tissue classification in phantom andanimal studies [37, 41]. Given these observations, micro-vibrations of the tissuemicrostructure caused by acoustic radiation force is another likely phenomenon.54The acoustic radiation force is related to both the acoustic energy and to the at-tenuation and scattering properties of the tissue, which are different for differentpathological tissue types.Despite the increased use of ultrasound RF time series as a method for tissueclassification, limited work has been done to provide evidence on the physical basisof tissue typing using this method. In the past, it has been shown that changing thetransmit power and the frame rate improves the tissue classification performanceof RF time series [32, 37, 41]. However, further experiments exploring the effectsof other ultrasound imaging parameters, which can increase the energy deliveredto the tissue, could provide sound evidence on the source of the tissue typing infor-mation extracted using RF time series.In this chapter, we build upon previous hypotheses by studying the effects ofadditional ultrasound imaging parameters such as imaging depth and time serieslength on tissue classification. We describe experimental studies on animal tissueto provide physical evidence of tissue typing using ultrasound RF time series. Theeffects of three imaging parameters: imaging depth, frame rate, and the length ofthe time series, on tissue classification are studied. In this work, RF time seriesdata from two animal tissues, bovine and chicken were collected and the effectsof these imaging parameters on tissue classification were studied. 13 experimentswere conducted and during each experiment a parameter was changed and the RFtime series data was saved. The data for each experiment was analysed by calculat-ing the RF time series features and then using two linear classifiers, linear SVM andlogistic regression for tissue classification. For each experimental dataset, the tis-sue classification performance was evaluated by a 20 fold stratified cross-validationand a calculation of cumulative area under the ROC curve.4.2 Materials and MethodsThis section describes the overall framework of this study. RF time series data wascollected for bovine and chicken. After collecting the data, we extracted seven RFtime series features for 200 1 mm2 ROIs centered around the ultrasound beam focalpoint. This resulted in 100 chicken and 100 beef (10 mm×10 mm window) trainingsamples described by seven RF time series features. For classification of the ROI55samples we used a logistic regression and a linear SVM classifier. The choiceof a simple linear approach to classification was made to minimize the effects ofmachine-learning model selection on the outcome of these comparative studies.The entire process was repeated 13 times with each experiment focusing on oneparameter, as described in Table 4.1. Figure 4.1 shows the experimental setup forthe study.Table 4.1: Parameters changed during each experiment. The parameters inbold were changed for the corresponding experiment.Exp. Frequency Focal Imaging Frame Time Series# (MHz) Point (cm) Depth (cm) Rate (fps) (frames)1 6.6 1.5 4 50 10002 6.6 1 4 50 10003 6.6 2 4 50 10004 6.6 2.5 4 50 10005 6.6 3 4 50 10006 6.6 3.5 4 50 10007 6.6 1.5 5.5 40 10008 6.6 1.5 3 60 10009 6.6 1.5 3 71 100010 6.6 1.5 4 50 20011 6.6 1.5 4 50 40012 6.6 1.5 4 50 60013 6.6 1.5 4 50 8004.2.1 Types of TissueTwo types of animal tissue were used, bovine tissue and chicken breast. Both tis-sues have fibrous structures which makes them harder to differentiate based on theB-mode appearance. Shown in Figure 4.2, the B-mode appearance of the two tissuetypes is very similar. This lack of visually distinct appearance of the tissue typesin B-mode images suggests a need for an ultrasound based tissue differentiationmethod that is independent of B-mode appearance. Ultrasound RF time series is atissue typing method that does not rely on B-mode appearance but in this methodof analysis spectral analysis of ultrasound time series signals is performed. The56Figure 4.1: Experiment setup. The two animal tissues are Chicken breast onthe left and bovine on the right surrounded by an ultrasound absorbingpad.animal tissues were bought fresh from the butcher shop 60 minutes before the datacollection. This allowed the tissue samples to attain room temperature. The twotypes of tissue were separated by ultrasound gel and imaged together, as shown inFigure 4.1. The ultrasound B-mode image of the two tissue types can be seen inFigure 4.2.4.2.2 Data CollectionData for this study was collected on a SonixMDP ultrasound machine using a L14-3/38 linear ultrasound probe at Ultrasonix Medical Corp., Richmond, BC, Canada.The two pieces of animal tissue were placed together in a tub surrounded byan absorbing pad. Ultrasound gel was put between the two tissues to ensure goodcontact. To minimize operator hand motion and collect data from the same tissuelocation for all experiments the ultrasound probe was held in place by a clamp. Be-57Figure 4.2: B-mode image of the two animal tissues. Bovine on the left andchicken (right).tween each experiment, we waited four minutes to minimize the effect of previousround of ultrasound RF data collection in terms of heating. Figure 4.1 shows theexperimental setup for the conducted experiments.4.2.3 Experiment Design13 experiments were conducted. For each experiment one ultrasound imaging pa-rameter was changed and the RF data was stored. The parameter changed for each58experiment is shown in Table 4.1. The default imaging parameters were as follows:Center frequency = 6.6 MHz, focal point = 1.5 cm, imaging depth = 4 cm, framerate = 50, and RF time series length = 1000 frames.4.2.4 Features100 ROIs of size 1 mm2 centered at the ultrasound beam focal point were extractedfor each tissue type. Six RF time series features and the fractal dimension featurewas calculated for each ROI. The RF frame size was 256×2080. The number ofpixels in 1 mm2 ROIs was dependent on the imaging depth. The ROI sizes in theRF domain for the each imaging depth, can be found in Table 4.2. Description ofthe seven RF time series features can be found in Section 2.2.2.Table 4.2: 1 mm2 ROI size in the RF domain for different imaging depths.Imaging depth (cm) 1 mm2 ROI size (RF samples)3 6×694 6×525.5 6×384.2.5 Classification and Cross-validationClassification: In this study, a logistic regression classifier and a linear SVM (nokernel trick) classifier were used. We chose simple linear classifiers to avoid over-fitting and show the robustness of the RF time series method. The results of thereport can further be improved by using state-of-the-art classifiers like kernal SVM(kernel machines) and Random Forests. However, in this chapter the goal is toobserve trends in response to varying parameters of ultrasound imaging. We usea simple model to minimize the effects of cross-validation on the outcome. Twoclassifiers were used to prove the stability of the results. The performance of eachclassifier was measured by computing the AUCs. We also calculated the p-valuefor statistically comparing the performance of the two classifiers used in this work.Cross-validation: A stratified k-fold cross-validation technique was employedto evaluate the classification performance. 20 stratified (equal number of test sam-59ples from each class) training and test sets were formed and a combined ROC curvewas computed for the 20 folds.4.3 ResultsTable 4.3 Table 4.4 Table 4.5 shows the areas under the ROC curve for the 13 ex-periments performed. As can be seen in the figures, the AUC values for both, linearSVM and logistic regression, are very similar. The classification performance dif-ference between the linear SVM classifier and the logistic regression classifier wasstatistically not significant (p > 0.05). The statistical insignificance (p > 0.05) inthe performance of the two classification algorithms proves the stability of the re-sults. AUC values using an RBF kernel SVM are also reported. The effects of theimaging depth, frame rate, and time series length on tissue typing are presented inthe following sections.4.3.1 Effects of Imaging Depth on Tissue ClassificationAfter analysis of the AUC values and performing statistical significance tests, itis apparent that increasing the ultrasound imaging depth significantly (p < 0.001)decreased the tissue classification performance. This decreasing trend can be seenin Figure 4.3. The difference in AUC values for different imaging depths (experi-ments 2, 3, 4, 5 and 6) was statistically highly significant (p-value = 0.000005) andcan be found in Table 4.3.Table 4.3: Area under the ROC curve with changing imaging depth.Area under ROC curve ParameterExp. # Linear Logistic RBF ImagingSVM Regression SVM Depth (cm)2 71.8% 73.8% 79.9% 13 88.7% 88.4% 86.2% 24 82.9% 82.6% 82.5% 2.55 79.1% 79.0% 77.3% 36 69.6% 69.0% 69.5% 3.560Figure 4.3: The effect of imaging depth on the classification accuracy of thetwo tissue types, bovine and chicken. As can be seen in the figure,increasing the imaging depth decreased the classification accuracy. Inother words, using the RF time series method, targets that are closer tothe ultrasound transducer can be classified more accurately than the tar-gets that are farther away from the transducer. The results are consistentwith the previous studies.4.3.2 Effects of Frame Rate on Tissue ClassificationThe AUC values for the experiments performed to study the effects of ultrasoundframe rate on tissue classification are reported in Table 4.4. From the AUC values,it is apparent that increasing the frame rate resulted in improved tissue classifica-tion. This improvement in the classification results as a result of increasing framerate was statistically significant (p-value = 0.027). The increasing AUC trend as aresult of increasing frame rate can be seen in Figure 4.4.61Table 4.4: Area under the ROC curve for different frame rates.Area under ROC curve ParameterExp. # Linear Logistic RBF Frame RateSVM Regression SVM (fps)7 77.7% 78.5% 77.2% 401 79.7% 79.3% 79.5% 508 81.2% 82.6% 84.2% 609 87.7% 87.1% 92.6% 71Figure 4.4: The effect of increasing frame rate on the classification accuracyof the two tissue types, bovine and chicken. Increasing the frame ratesignificantly improved the classification results. It is apparent from theabove figure, increasing the frame rate from 40 to 71 frames per sec-ond improved the area under the ROC curve from 77% to 87%. Pre-vious studies have shown that increasing the frame rate improved theclassification performance, therefore our results are consistent with theprevious studies.624.3.3 Effects of Time Series Length on Tissue ClassificationTable 4.5 shows the tissue classification performance in terms of the AUC valuesfor different time series lengths. It was found that increasing the time series lengthsignificantly (p-value = 0.00014) improved the tissue classification results. Thisobservation suggests that using more frames for RF time series feature calcula-tion can result in strong features that distinguish the tissue types more accurately.The increasing trend in the AUC values as the number of RF frames analyzed isincreased can be seen in Figure 4.5.Table 4.5: Area under the ROC curve for different time series length.Area under ROC curve ParameterExp. # Linear Logistic RBF Time SeriesSVM Regression SVM Length (frames)10 49.0% 43.5% 61.5% 20011 69.1% 68.8% 70.0% 40012 75.3% 75.8% 75.0% 60013 79.1% 78.1% 79.2% 8001 79.7% 79.3% 79.5% 1000To further investigate the decreased performance when using 200 frames (4seconds of RF data), we used the last 200 frames, instead of the first 200, andrecorded the AUC as 54% and 56% using Logistic Regressions and SVM classi-fiers. We also calculated the AUC using 200 RF frames and at an imaging depthof 2 cms (Experiment # 3), as compared to 1.5 cms, and the result was 52% forclassification using Logistic Regression and 53% using linear SVM.4.4 Discussions and ConclusionsThis chapter described an experimental study on animal tissue to empirically testRF time series as a method for tissue typing. Previous studies on ultrasound RFtime series and Chapter 2 and Chapter 3 have reported consistent tissue classifica-tion results using this method. The source of tissue typing information extractedfrom this type of analysis has not been empirically confirmed. In this chapter westudied the effects of ultrasound imaging depth, frame rate, and the length of the63Figure 4.5: The effect of the length of the time series on the classificationaccuracy of the two tissue types, bovine and chicken. As can be seenin the above figure, increasing the number of analysed RF frames, forthe RF time series feature calculation, improved the classification per-formance. These results show, the tissue typing information improvedas more RF frames are analysed. These observations are consistent withthe previous published literature.time series on the classification of two animal tissue types. It has been demon-strated that increasing the energy delivered to the tissue improved the tissue classi-fication accuracy.The results of this study suggest that:1. Using ultrasound RF time series, as the ultrasound imaging depth increases,the animal tissue classification performance significantly decreases (p< 0.001).2. Increasing the frame rate of the ultrasound machine significantly improvedthe animal tissue classification result using ultrasound RF time series method(p < 0.05).643. In ultrasound RF time series analysis, increasing the number of RF framesanalyzed (time series length) significantly enhances the animal tissue classi-fication performance (p < 0.001).A lower tissue classification accuracy was observed at an imaging depth of 1cm. This could be attributed to the near field behaviour of the ultrasound beamand to the fact that 1 cm was below the focal range of the transducer. In ourexperiments, using four seconds of data at 50 frames per second resulted in poorclassification results. On the other hand, as more frames were analysed, the tissueclassification was significantly better. This could be due to the temperature increasein the tissue as a result of the increased time series length.Overall, it is clear that increasing the amount of the energy delivered to the tis-sue improves the performance of the tissue classifiers. The improved classificationperformance as a result of changing the imaging depth, frame rate, and the lengthof the time series could be due to the the temperature increase in controlled irra-diation of tissue with RF time series [87]. Increased energy delivered to the tissuethrough increased frame rate and shallower imaging depth could also be the causeof the improved tissue classification accuracy [37, 41]. An additional probable phe-nomenon for the improved tissue classification result could be micro-vibrations ofthe tissue microstructure caused by acoustic radiation force. The acoustic radiationforce is related to both the acoustic energy and to the attenuation and scatteringproperties of the tissue, which are different for different pathological tissue types.Simulation studies to measure these micro-vibrations are ongoing to further vali-date these ultrasound RF time series hypotheses.The results of this study are consistent with the previous reported results whichsuggest that increased energy delivered to the tissue through increased frame rateor transmit power results in improved tissue classification accuracy in phantom andanimal studies [37, 41].4.5 Chapter SummaryIn this chapter the effects of three ultrasound imaging parameters–imaging depth,frame rate, and RF time series length–on tissue classification performance werestudied to provide further evidence on the source of tissue typing information ex-65tracted using ultrasound RF time series. The results of this chapter are consistentwith the previously reported studies, which suggest that increased energy deliv-ered to the tissue results in improved tissue classification. From the observationsreported in this chapter, it can be concluded that decreasing the imaging depth,increasing the frame rate, and increasing the length of the time series improvesanimal tissue classification using ultrasound RF time series.66Chapter 5ConclusionsAugmentation of the current breast and prostate cancer diagnosis is a pressing needdue to the rate of over-diagnosis and number of unnecessary biopsies performed. Inthis work, an ultrasound-based tissue classification method for breast and prostatecancer diagnosis was presented.Ultrasound is a rapidly growing imaging modality that is inexpensive as com-pared to MRI and unlike X-ray or CT, does not expose the patient to ionizingradiation. Ultrasound-based cancer diagnosis methods are versatile and do not re-quire additional hardware. Ultrasound RF time series is an ultrasound-based cancerdiagnosis technique that is non invasive and has been proven successful in classi-fication of animal tissue [37], prostate cancer ex vivo [32] and in vivo [39, 40],and tissue abalation monitoring [41]. Ultrasound RF time series along with otherultrasound-based diagnosis methods could provide radiologists a real-time breastand prostate cancer diagnosis tool that could significantly reduce the number ofnegative biopsies.A tissue classification framework, comprising ultrasound RF time series sig-nal analysis, ultrasound RF signal analysis, B-mode texture analysis, and machinelearning techniques has been proposed to address the need for ultrasound-basedcancer diagnosis. Breast and prostate tissue classification studies were performedto evaluate the performance of the tissue classification framework and prove thevalidity of RF time series analysis as a tissue typing method. The results of thesestudies suggest the potential of ultrasound RF time series as a useful cancer di-67agnosis method. Cancer probability maps presented in this work could provideradiologists with key information about the cancer affected region, which couldimprove biopsy targeting and in some cases eliminating the need for biopsy.5.1 Summary of Contributions• We developed a method for classifying breast lesions based on ultrasound RFtime series analysis. This was the first RF time series study on breast data.Data from 22 subjects was analyzed and promising results were presented.Ultrasound RF time series, single RF frame, and B-mode texture featureswere used along with a machine learning framework to diagnose cancer inbreast lesions. The clinical significance of this work are the reported cancerprobability maps. Cancer probability maps could provide radiologist a real-time cancer diagnosis tool which could significantly improve the cancer yieldand reduce the number of unnecessary breast biopsies. Feature calculationscripts were developed to calculate RF time series features, B-mode texturefeatures, and RF spectral features. A machine learning package was alsodeveloped for classification of ROIs described by the calculated features.Data analysis and visualization scripts were also developed for exploratoryanalysis.• A new semi-supervised machine learning technique was proposed for in vivoprostate tissue classification. The new semi-supervised classification methodperforms clustering of the data prior to classification to eliminate the effectof outliers and improve the differentiation between the malignant and thebenign subjects. The effect of outliers was minimized by training classi-fier models on cluster specific samples. In this prostate tissue classifica-tion study, ultrasound RF time series features were used for classificationof MRI-targeted biopsy cores. Data from 18 cores (14 subjects) was anal-ysed and highly accurate cancer prediction was reported. The results of thisstudy suggests that clustering of data before classification can improve theclassification results. This work could potentially lead to the fusion of theproposed tissue classification framework with the Philips UroNav platformto complement MRI-targeted TRUS guided prostate biopsy and provide real-68time diagnosis of prostate cancer. Feature calculation scripts were developedin MATLAB and the machine learning algorithms were written in Python.The developed algorithms could benefit future studies on ultrasound RF timeseries.• An animal tissue study was also completed to substantiate the ultrasoundRF time series hypothesis. The effects of ultrasound imaging parameters:imaging depth, frame rate, and the time series length were studied and theresults were documented. This work could provide more evidence on thephysical basis of tissue typing using ultrasound RF time series method. Datafrom two animal tissue, steak and chicken, was collected and analysed. Ul-trasound RF time series features were calculated and linear classifiers wereemployed for classification. The results of this study are consistent with thepreviously reported and simulation studies are ongoing to further validatethe RF time series hypotheses. Feature calculation scripts along with ma-chine learning algorithms were developed for this study. A comprehensiveanalysis of the results was performed to study the effect of each ultrasoundimaging parameter on the tissue classification accuracy.5.2 Future WorkFuture work in this field would include researching more robust features for tissueclassification. Wavelet analysis of RF time series along with the current featurescould result in improved tissue classification.The methods presented in this thesis may be incorporated into the PhilipsUroNav platform consequently providing radiologists with a real-time cancer di-agnosis tool.Experimental studies further exploring the source of tissue typing informationusing RF time series would prove beneficial to future developments employing theRF time series technique. Simulations studies to measure the micro-vibrations andtemperature changes in the tissue due to energy delivered would help prove thephysical basis of tissue typing using RF time series.Hyper-parameter optimization techniques would significantly reduce the com-putation time required to select the optimum parameters for a classification algo-69rithm. Currently, a grid search over the parameter space is performed. This ap-proach is time consuming and not practical. It has been shown that random searchfor hyper-parameter optimization is more efficient than the traditional grid searchor manual search for finding the best hyper-parameter of a classification algorithm[89]. For real-time cancer diagnosis the computation time would have to be sig-nificantly reduced and this can be achieved by developing or employing advancedhyper-parameter optimization techniques like random search.Approaches such as outlier detection, Density-based spatial clustering of ap-plications with noise (DBSCAN) [90], and label propagation [91, 92] are some ofthe techniques that could improve the current classification approach. By using ad-vanced clustering algorithms like DBSCAN, which is more robust to noisy dataset,the novel semi-supervised machine learning approach presented in Chapter 3 couldbe improved.70Bibliography[1] Siegel, R., Ma, J., Zou, Z., Jemal, A.: Cancer statistics, 2014. CA: A CancerJournal for Clinicians 64(1) (2014) 9–29 → pages 1[2] Jemal, A., Bray, F., Center, M.: Global cancer statistics. CA: A CancerJournal for Clinicians 61(2) (2011) 69–90 → pages 1[3] Poplack, S., Carney, P., Weiss, J.: Screening Mammography: Costs and Useof Screening-related Services1. Radiology 234(1) (2005) 79–85 → pages 1,7[4] Stavros, A., Thickman, D., Rapp, C.: Solid breast nodules: use ofsonography to distinguish between benign and malignant lesions. Radiology196(1) (1995) 123–134 → pages 1, 2, 8[5] Miller, A., Wall, C., Baines, C., Sun, P.: Twenty five year follow-up forbreast cancer incidence and mortality of the Canadian National BreastScreening Study: randomised screening trial. BMJ: British Medical Journal348 (2014) 1–10 → pages 2[6] Mandelson, M.T., Porter, P.L., White, D., Finder, C.A., Taplin, S.H., White,E.: Breast density as a predictor of mammographic detection: comparison ofinterval-and screen-detected cancers. Journal of the National CancerInstitute 92(13) (2000) 1081–1087 → pages 2[7] Boyd, N., Guo, H., Martin, L.: Mammographic density and the risk anddetection of breast cancer. New England Journal of Medicine 356(3) (2007)227–236 → pages 2, 8[8] Kolb, T., Lichy, J., Newhouse, J.: Comparison of the performance ofscreening mammography, physical examination, and breast us andevaluation of factors that influence them: an analysis of 27,825. Radiology225(1) (2002) 165–175 → pages 2, 7, 871[9] Berg, W., Blume, J.: Combined screening with ultrasound andmammography vs mammography alone in women at elevated risk of breastcancer. JAMA: the journal of the American Medical Association 299(18)(2008) 2151–2163 → pages[10] Kelly, K.M., Dean, J., Comulada, W.S., Lee, S.J.: Breast cancer detectionusing automated whole breast ultrasound and mammography inradiographically dense breasts. European Radiology 20(3) (2010) 734–42→ pages 2, 8[11] Rahbar, G., Sie, A., Hansen, G.: Benign versus Malignant Solid BreastMasses: US Differentiation1. Radiology 213(3) (1999) 889–894 → pages 2,8[12] Cheng, H., Shan, J., Ju, W., Guo, Y., Zhang, L.: Automated breast cancerdetection and classification using ultrasound images: A survey. PatternRecognition 43(1) (2010) 299–317 → pages 2, 9[13] Alam, S., Feleppa, E., Rondeau, M.: Ultrasonic multi-feature analysisprocedure for computer-aided diagnosis of solid breast lesions. Ultrasonicimaging 38(1) (2011) 17–38 → pages 8[14] Burnside, E., Hall, T., Sommer, A.: Differentiating benign from malignantsolid breast masses with US strain Imaging1. Radiology 245(2) (2007)401–410 → pages 9[15] Berg, W., Cosgrove, D., Dore´, C.: Shear-wave elastography improves thespecificity of breast US: the BE1 multinational study of 939 masses.Radiology 262(2) (2012) 435–449 → pages 9[16] Doherty, J.R., Trahey, G.E., Nightingale, K.R., Palmeri, M.L.: Acousticradiation force elasticity imaging in diagnostic ultrasound. IEEETransactions on Ultrasonics, Ferroelectrics and Frequency Control 60(4)(2013) 685–701 → pages 9[17] Tan, T., Platel, B., Mus, R., Tabar, L., Mann, R.M., Karssemeijer, N.:Computer-aided detection of cancer in automated 3-D breast ultrasound.IEEE Transactions on Medical Imaging 32(9) (2013) 1698–706 → pages 2,9[18] Margel, D., Yap, S.a., Lawrentschuk, N., Klotz, L., Haider, M., Hersey, K.,Finelli, A., Zlotta, A., Trachtenberg, J., Fleshner, N.: Impact ofmultiparametric endorectal coil prostate magnetic resonance imaging on72disease reclassification among active surveillance candidates: a prospectivecohort study. The Journal of urology 187(4) (2012) 1247–52 → pages 2, 36[19] Hambrock, T., Vos, P.C., Hulsbergen–van de Kaa, Christina A Barentsz,J.O., Huisman, H.J.: Prostate cancer: computer-aided diagnosis withmultiparametric 3-T MR imagingeffect on observer performance. Radiology266(2) (2013) 521–530 → pages[20] Rooij, M., Hamoen, E.H.J., Fu¨tterer, J.J., Barentsz, J.O., Rovers, M.M.:Accuracy of multiparametric MRI for prostate cancer detection: ameta-analysis. American Journal of Roentgenology 202(2) (2014) 343–51→ pages[21] Moradi, M., Salcudean, S.E., Chang, S.D., Jones, E.C., Buchan, N., Casey,R.G., Goldenberg, S.L., Kozlowski, P.: Multiparametric MRI maps fordetection and grading of dominant prostate tumors. Journal of MagneticResonance Imaging 35(6) (2012) 1403–1413 → pages 2[22] Pinto, P., Chung, P., Rastinehad, A.: Ultrasound fusion guided prostatebiopsy improves cancer detection following transrectal ultrasound biopsyand correlates with multiparametric magnetic resonance. The Journal ofUrology 186(4) (2011) 1281–1285 → pages 2, 36, 37[23] Barentsz, J.O., Richenberg, J., Clements, R., Choyke, P., Verma, S., Villeirs,G., Rouviere, O., Logager, V., Fu¨tterer, J.J.: ESUR prostate MR guidelines2012. European Radiology 22(4) (2012) 746–57 → pages 2, 36[24] Natarajan, S., Marks, L.S., Margolis, D.J.a., Huang, J., Macairan, M.L.,Lieu, P., Fenster, A.: Clinical application of a 3D ultrasound-guided prostatebiopsy system. Urologic Oncology 29(3) (2011) 334–42 → pages 2, 36[25] Bax, J., Smith, D., Bartha, L., Montreuil, J., Sherebrin, S., Gardi, L.,Edirisinghe, C., Fenster, A.: A compact mechatronic system for 3Dultrasound guided prostate interventions. Medical Physics 38(2) (2011)1055 → pages 2, 36[26] Xu, S., Kruecker, J., Turkbey, B.: Real-time MRI-TRUS fusion for guidanceof targeted prostate biopsies. Computer Aided Surgery 13(5) (2008)255–264 → pages 2, 36[27] Feleppa, E., Porter, C., Ketterling, J.: Recent Advances in UltrasonicTissue-Type Imaging of the Prostate (2007) → pages 2, 3673[28] Pallwein, L., Mitterberger, M., Struve, P., Pinggera, G., Horninger, W.,Bartsch, G., Aigner, F., Lorenz, A., Pedross, F., Frauscher, F.: Real-timeelastography for detecting prostate cancer: preliminary experience. BJUIternational 100(1) (2007) 42–46 → pages 2, 36[29] Moradi, M., Mahdavi, S.S., Nir, G., Mohareri, O., Koupparis, A., Gagnon,L.O., Fazli, L., Casey, R.G., Ischia, J., Jones, E.C., et al.: Multiparametric3d in vivo ultrasound vibroelastography imaging of prostate cancer:Preliminary results. Medical Physics 41(7) (2014) 073505 → pages 2[30] Scheipers, U., Ermert, H., Sommerfeld, H.J., Garcia-Schu¨rmann, M., Senge,T., Philippou, S.: Ultrasonic multifeature tissue characterization for prostatediagnostics. Ultrasound in Medicine & Biology 29(8) (2003) 1137–1149 →pages 2, 36[31] Moradi, M., Mousavi, P., Abolmaesumi, P.: Computer-aided diagnosis ofprostate cancer with emphasis on ultrasound-based approaches: a review.Ultrasound in Medicine & Biology 33(7) (2007) 1010–1028 → pages 2[32] Moradi, M., Abolmaesumi, P., Siemens, D.R., Sauerbrei, E.E., Boag, A.H.,Mousavi, P.: Augmenting detection of prostate cancer in transrectalultrasound images using SVM and RF time series. IEEE Transactions onBiomedical Engineering 56(9) (2009) 2214–2224 → pages 3, 39, 54, 55, 67[33] Feleppa, E.J., Kalisz, A., Sokil-melgar, J.B., Lizzi, F.L., Liu, T., Rosado,A.L., Shao, M.C., Fair, W.R., Wang, Y., Cookson, M.S., Reuter, V.E.,Heston, W.D.W.: Typing of prostate tissue by ultrasonic spectrum analysis.IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency Control43(4) (1996) 609–619 → pages 3, 10[34] Moradi, M., Abolmaesumi, P., Isotalo, P., Siemens, D., Sauerbrei, E.,Mousavi, P.: Detection of prostate cancer from RF ultrasound echo signalsusing fractal analysis. In: Proceedings of 28th Annual InternationalConference of the IEEE Engineering in Medicine and Biology Society.(2006) 2400 –2403 → pages 3, 14[35] Moradi, M., Mousavi, P., Siemens, D.R., Sauerbrei, E.E., Isotalo, P., Boag,A., Abolmaesumi, P.: Discrete Fourier analysis of ultrasound RF time seriesfor detection of prostate cancer. In: Proceedings of 29th InternationalConference of the IEEE Engineering in Medicine and Biology Society.(2007) 1339–1342 → pages 3, 1474[36] Moradi, M., Mousavi, P., Abolmaesumi, P.: Tissue characterization usingfractal dimension of high frequency ultrasound RF time series. In: MICCAI.Volume 4792 of LNCS. (2007) 900–908 → pages 3, 36[37] Moradi, M., Abolmaesumi, P., Mousavi, P.: Tissue typing using ultrasoundRF time series: Experiments with animal tissue samples. Medical Physics37(8) (2010) 4401–4413 → pages 3, 4, 33, 54, 55, 65, 67[38] Imani, F., Daoud, M., Moradi, M., Abolmaesumi, P., Mousavi, P.: Tissueclassification using depth-dependent ultrasound time series analysis: in-vitroanimal study. In: SPIE Medical Imaging. Volume 7968., InternationalSociety for Optics and Photonics (2011) 79680F–79680F–7 → pages 3[39] Moradi, M., Mahdavi, S.S., Nir, G., Jones, E.C., Goldenberg, S.L.,Salcudean, S.E.: Ultrasound rf time series for tissue typing: first in vivoclinical results. In: SPIE Medical Imaging, International Society for Opticsand Photonics (2013) 86701I–86701I → pages 4, 54, 67[40] Imani, F., Abolmaesumi, P., Gibson, E.: Ultrasound-based characterizationof prostate cancer: an in vivo clinical feasibility study. In: MICCAI. Volume8150 of LNCS. (2013) 279–286 → pages 4, 39, 54, 67[41] Imani, F., Abolmaesumi, P., Wu, M.Z., Lasso, A., Burdette, E.C., Ghoshal,G., Heffter, T., Williams, E., Neubauer, P., Fichtinger, G., Mousavi, P.:Ultrasound-guided characterization of interstitial ablated tissue using RFtime series: feasibility study. IEEE Transactions on Bio-medicalEngineering 60(6) (2013) 1608–18 → pages 4, 54, 55, 65, 67[42] Siegel, R.: Cancer statistics, 2013. CA: A Cancer Journal for Clinicians63(1) (2013) 11–30 → pages 7[43] Boyle, P., Levin, B.: World cancer report 2008. IARC Press, InternationalAgency for Research on Cancer (2008) → pages 7[44] Stomper, P., D’souza, D.: Analysis of parenchymal density onmammograms in 1353 women 25-79 years old. AJR. American Journal ofRoentgenology 167(5) (1996) 1261–1265 → pages 8[45] Berg, W.A., Gutierrez, L., Nessaiver, M.S., Carter, W.B., Bhargavan, M.,Lewis, R.S., Ioffe, O.B.: Diagnostic accuracy of mammography, clinicalexamination, US, and MR imaging in preoperative assessment of breastcancer. Radiology 233(3) (2004) 830–849 → pages 875[46] Zheng, Y., Greenleaf, J.F., Gisvold, J.J.: Reduction of breast biopsies with amodified self-organizing map. IEEE Transactions on Neural Networks 8(6)(1997) 1386–96 → pages 8[47] Chen, D., Chang, R., Huang, Y.: Computer-aided Diagnosis Applied to USof Solid Breast Nodules by Using Neural Networks. Radiology 213(2)(1999) 407–412 → pages 8, 13[48] Giger, M.L., Al-Hallaq, H., Huo, Z., Moran, C., Wolverton, D.E., Chan,C.W., Zhong, W.: Computerized analysis of lesions in US images of thebreast. Academic Radiology 6(11) (1999) 665–74 → pages 8[49] Lefebvre, F., Meunier, M., Thibault, F.: Computerized ultrasound B-scancharacterization of breast nodules. Ultrasound in Medicine & Biology 26(9)(2000) 1421–1428 → pages 8[50] Donohue, K., Huang, L., Burks, T., Forsberg, F., Piccoli, C.: Tissueclassification with generalized spectrum parameters. Ultrasound in Medicine& Biology 27(11) (2001) 1505–1514 → pages 8, 13[51] Shankar, P.M., Dumane, V.a., Reid, J.M., Genis, V., Forsberg, F., Piccoli,C.W., Goldberg, B.B.: Classification of ultrasonic B-mode images of breastmasses using Nakagami distribution. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control 48(2) (2001) 569–80 → pages 8[52] Shankar, P.M., Dumane, V.a., Piccoli, C.W., Reid, J.M., Forsberg, F.,Goldberg, B.B.: Computer-aided classification of breast masses in ultrasonicB-scans using a multiparameter approach. IEEE Transactions onUltrasonics, Ferroelectrics and Frequency Control 50(8) (2003) 1002–9 →pages 8[53] Chang, R.F., Wu, W.J., Moon, W.K., Chou, Y.H., Chen, D.R.: Supportvector machines for diagnosis of breast tumors on US images. AcademicRadiology 10(2) (2003) 189–197 → pages 8[54] Chang, R.F., Wu, W.J., Moon, W.K., Chen, D.R.: Improvement in breasttumor discrimination by support vector machines and speckle-emphasistexture analysis. Ultrasound in Medicine & Biology 29(5) (2003) 679–686→ pages 8[55] Huang, Y.L., Chen, D.R.: Support vector machines in sonography:application to decision making in the diagnosis of breast cancer. ClinicalImaging 29(3) (2005) 179–84 → pages 876[56] Drukker, K., Giger, M.L., Mendelson, E.B.: Computerized analysis ofshadowing on breast ultrasound for improved lesion detection. MedicalPhysics 30(7) (2003) 1833 → pages 8[57] Drukker, K., Giger, M.L., Metz, C.E.: Robustness of computerized lesiondetection and classification scheme across different breast US platforms.Radiology 237(3) (2005) 834–840 → pages 8[58] Joo, S., Yang, Y.S., Moon, W.K., Kim, H.C., Abstract, A.: Computer-aideddiagnosis of solid breast nodules: use of an artificial neural network basedon multiple sonographic features. IEEE Transactions on Medical Imaging23(10) (2004) 1292–1300 → pages 8[59] Chen, D.R., Chang, R.F., Chen, C.J., Ho, M.F., Kuo, S.J., Chen, S.T., Hung,S.J., Moon, W.K.: Classification of breast ultrasound images using fractalfeature. Clinical imaging 29(4) (2005) 235–245 → pages 8[60] Gomez, W., Pereira, W.C.a., Infantosi, a.F.C.: Analysis of co-occurrencetexture statistics as a function of gray-level quantization for classifyingbreast ultrasound. IEEE Transactions on Medical Imaging 31(10) (2012)1889–99 → pages 8, 13[61] Yang, M.C., Moon, W.K., Wang, Y.C., Bae, M.S., Huang, C.S., Chen, J.H.,Chang, R.F.: Robust Texture Analysis Using Multi-resolution Gray-scaleInvariant Features for Breast Sonographic Tumor Diagnosis. IEEETransactions on Medical Imaging 32(12) (2013) 2262–2273 → pages 9, 13[62] Garra, B.S., Cespedes, E.I., Ophir, J., Spratt, S.R., Zuurbier, R.A., Magnant,C.M., Pennanen, M.F.: Elastography of breast lesions: initial clinical results.Radiology 202(1) (1997) 79–86 → pages 9[63] Bercoff, J., Tanter, M., Fink, M.: Supersonic shear imaging: a new techniquefor soft tissue elasticity mapping. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control 51(4) (2004) 396–409 → pages 9[64] Tanter, M., Bercoff, J., Athanasiou, A., Deffieux, T., Gennisson, J.L.,Montaldo, G., Muller, M., Tardivon, A., Fink, M.: Quantitative assessmentof breast lesion viscoelasticity: initial clinical results using supersonic shearimaging. Ultrasound in Medicine & Biology 34(9) (2008) 1373–1386 →pages[65] Athanasiou, A., Tardivon, A., Tanter, M.: Breast lesions: quantitativeelastography with supersonic shear imaging–preliminary results. Radiology256(1) (2010) 297–303 → pages 977[66] Uniyal, N., Eskandari, H., Abolmaesumi, P., Sojoudi, S., Gordon, P.,Warren, L., Rohling, R.N., Salcudean, S.E., Moradi, M.: A new approach toultrasonic detection of malignant breast tumors. In: Ultrasonics Symposium(IUS), 2013 IEEE International. (July 2013) 96–99 → pages 9, 54[67] Eskandari, H., Salcudean, S.E., Rohling, R., Baghani, A., Frew, S., Gordon,P.B., Warren, L.: Identifying Malignant and Benign Breast Lesions UsingVibroelastography. In: Ultrasonics Symposium (IUS), 2013 IEEEInternational. (2013) 25–28 → pages 10[68] Zahiri-Azar, R., Salcudean, S.E.: Motion estimation in ultrasound imagesusing time domain cross correlation with prior estimates. IEEE Transactionson Biomedical Engineering 53(10) (2006) 1990–2000 → pages 11, 33[69] Garra, B., Krasner, B., Horii, S.: Improving the distinction between benignand malignant breast lesions: the value of sonographic texture analysis.Ultrasonic Imaging 15(4) (1993) 267–285 → pages 13[70] Horsch, K., Giger, M.L., Venta, L.a., Vyborny, C.J.: Computerized diagnosisof breast lesions on ultrasound. Medical Physics 29(2) (2002) 157 → pages[71] Chen, D., Chang, R., Kuo, W.: Diagnosis of breast tumors with sonographictexture analysis using wavelet transform and neural networks. Ultrasound inMedicine & Biology 28(10) (2002) 1301–1310 → pages 13[72] Higuchi, T.: Approach to an irregular time series on the basis of the fractaltheory. Physica D: Nonlinear Phenomena 30(2) (1988) 277—-283 → pages14[73] Lizzi, F., Ostromogilsky, M.: Relationship of ultrasonic spectral parametersto features of tissue microstructure. IEEE Transactions on Ultrasonics,Ferroelectrics and Frequency Control 34(3) (1987) 319–29 → pages 14[74] Lizzi, F., Feleppa, E., Alam, S.K., Deng, C.: Ultrasonic spectrum analysisfor tissue evaluation. Pattern Recognition Letters 24(4) (2003) 637–658 →pages 14[75] Bishop, C.M.: Pattern Recognition and Machine Learning. Springer Science(2006) → pages 17, 18[76] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel,O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J.,Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.:78Scikit-learn: Machine learning in Python. The Journal of Machine LearningResearch 12 (2011) 2825–2830 → pages 17, 19, 41[77] Platt, J.: Probabilistic outputs for support vector machines and comparisonsto regularized likelihood methods. Advances in Large Margin Classifiers10(3) (1999) 61–74 → pages 18, 41[78] Chang, C., Lin, C.: LIBSVM: a library for support vector machines. ACMTransactions on Intelligent Systems and Technology (TIST) 2(3) (2011)1–39 → pages 18[79] Breiman, L.: Random forests. Machine Learning 45(1) (2001) 5–32 →pages 18, 19[80] Criminisi, A., Shotton, J.: Decision Forests for Computer Vision andMedical Image Analysis. (2013) → pages 18, 19, 24[81] Hastie, T., Tibshirani, R., Friedman, J.J.H.: The elements of statisticallearning. Volume 1. Springer New York (2001) → pages 19[82] Guyon, I., Weston, J., Barnhill, S., Vapnik, V.: Gene selection for cancerclassification using support vector machines. Machine Learning 46(1-3)(2002) 389–422 → pages 20[83] Robin, X., Turck, N., Hainard, A., Tiberti, N., Lisacek, F., Sanchez, J.C.,Mu¨ller, M.: pROC: an open-source package for R and S+ to analyze andcompare ROC curves. BMC Bioinformatics 12(1) (January 2011) 77 →pages 21[84] Hanley, J., McNeil, B.: A method of comparing the areas under receiveroperating characteristic curves derived from the same cases. Radiology148(3) (1983) 839–843 → pages 21, 24[85] Chawla, N., Japkowicz, N., Kotcz, A.: Editorial: special issue on learningfrom imbalanced data sets. ACM SIGKDD Explorations Newsletter 6(1)(2004) 1–6 → pages 23[86] He, H., Garcia, E.: Learning from imbalanced data. IEEE Transactions onKnowledge and Data Engineering 21(9) (2009) 1263–1284 → pages 23[87] Daoud, M.I., Mousavi, P., Imani, F., Rohling, R., Abolmaesumi, P.: Tissueclassification using ultrasound-induced variations in acoustic backscatteringfeatures. IEEE Transactions on Biomedical Engineering 60(2) (2013)310–20 → pages 33, 54, 6579[88] Rapiti, E., Schaffar, R., Iselin, C., Miralbell, R., Pelte, M.F., Weber, D.,Zanetti, R., Neyroud-Caspar, I., Bouchardy, C.: Importance anddeterminants of Gleason score undergrading on biopsy sample of prostatecancer in a population-based study. BMC Urology 13(1) (2013) 19 →pages 35[89] Bergstra, J., Bengio, Y.: Random search for hyper-parameter optimization.The Journal of Machine Learning Research 13(1) (2012) 281–305 → pages70[90] Ester, M., Kriegel, H., Sander, J., Xu, X.: A density-based algorithm fordiscovering clusters in large spatial databases with noise. In: KDD. (1996)226–231 → pages 70[91] Bengio, Y., Delalleau, O., Le Roux, N.: Label propagation and quadraticcriterion. In: Semi-supervised learning. (2006) 193–216 → pages 70[92] Delalleau, O., Bengio, Y., Le Roux, N.: Efficient non-parametric functioninduction in semi-supervised learning. In: Proceedings of the TenthInternational Workshop on Artificial Intelligence and Statistics. (2005)96–103 → pages 7080