Lancaster-Dissertation-2018
Total Page:16
File Type:pdf, Size:1020Kb
c Copyright by Keith Lancaster 2018 All Rights Reserved ASYMMETRY MEASURES FOR AUTOMATED MELANOMA DETECTION IN DERMOSCOPIC IMAGES A Dissertation Presented to the Faculty of the Department of Electrical and Computer Engineering University of Houston In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy By Keith Lancaster May 2018 ASYMMETRY MEASURES FOR AUTOMATED MELANOMA DETECTION IN DERMOSCOPIC IMAGES Keith Lancaster Approved: Chair of the Committee Dr. George Zouridakis, Professor Electrical and Computer Engineering Committee Members: Dr. Ji Chen, Professor Electrical and Computer Engineering Dr. Ben H. Jansen, Professor Emeritus Electrical and Computer Engineering Dr. John Glover, Professor Emeritus Electrical and Computer Engineering Dr. Xiaojing Yuan, Associate Professor Computer Engineering Technology Dr. Suresh Khator, Associate Dean, Dr. Badri Roysam, Professor and Chair Cullen College of Engineering Electrical & Computer Engineering Acknowledgments First, I would like to thank Dr. George Zouridakis who has supported and en- couraged me throughout this long process. It was an honor to work with him, and I will be forever grateful. I would be remiss not to thank Dr. Ben Jansen, Dr. John Glover, Dr. Ji Chen, and Dr. Xiaojing Yuan for serving on my committee. I would also like to thank my lab mates Ning Situ, Tarun Wadhawan, and Rui Hu for their friendship and support. Finally, I must thank my wife Amy, son Garrett, and his wife, Cindy. Without their continued encouragement over the years, I would never have completed this journey. v ASYMMETRY MEASURES FOR AUTOMATED MELANOMA DETECTION IN DERMOSCOPIC IMAGES An Abstract of a Dissertation Presented to the Faculty of the Department of Electrical and Computer Engineering University of Houston In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy By Keith Lancaster May 2018 vi Abstract Dermoscopic rules such as the ABCD and Menzies rules are employed by der- matologists to determine the likelihood that a suspicious lesion is cancerous. This dissertation focuses on the improvement of automated melanoma recognition systems that implement these rules, specifically by enhancing the ability of these systems to recognize lesion asymmetry, a significant indicator of melanoma. Two approaches are proposed for asymmetry classification. The first utilizes the irregularity of the outer contour of the lesion combined with measures that com- pare quadrants of the lesion with respect to area, color, and melanin content. The second method uses size theory as the basis for determining asymmetry. In this ap- proach, measuring functions are employed to expose relevant characteristics of the lesion. The one-dimensional measuring functions are mapped into size functions in R2 and compared using the bottleneck distance. The distances are used as features for classification. Annotated dermoscopic images were used to train classifiers for both methods. Classification rates were competitive with other approaches for both methods inde- pendently, with the combined method exhibiting 95% accuracy. Additionally, decision fusion strategies were investigated as a means of combining the results from individual melanoma classifiers using the asymmetry methods devel- oped in this study. The best approach showed 100% sensitivity and 64% specificity, exceeding the performance of the individual classifiers. Finally, a software framework for the development of medical applications is pre- sented. This framework attempts to provide biomedical researchers with a simplified approach to creating mobile applications for medical processing. vii Table of Contents Acknowledgments ................................ v Abstract ...................................... vii Table of Contents ................................ viii List of Figures .................................. xi List of Tables ................................... xiii 1 Introduction .................................. 1 1.1 Research Objectives . 2 2 Background .................................. 5 2.1 Identifying Suspected Melanoma . 5 2.1.1 Diagnostic Methods Used by PCPs . 5 2.1.2 Diagnostic Methods Used by Dermatologists . 6 2.2 Dermoscopy . 6 2.2.1 Dermoscopic Rules . 7 2.3 Automating Melanoma Detection . 10 2.4 Classifier Fusion . 12 2.4.1 Fusion Topologies . 12 2.5 Utilizing Smartphones for Melanoma Screening . 14 2.6 Image Processing Frameworks for Mobile Devices . 15 viii 2.7 Related Works . 15 2.7.1 Automation of Dermoscopic Rules . 15 2.7.2 DermoScreen . 19 3 Methods .................................... 22 3.1 Data . 22 3.2 Process Overview . 23 3.2.1 Image Segmentation . 23 3.2.2 Feature Extraction . 23 3.2.3 Feature Selection . 24 3.2.4 Classification . 27 3.2.5 Assessing Classifier Performance . 30 3.3 Assessment of Lesion Asymmetry . 33 3.3.1 Introduction . 33 3.3.2 DermoScreen Asymmetry Algorithm . 33 3.3.3 Asymmetry Based on Size Theory . 35 3.3.4 Asymmetry from Border and Area Characteristics . 39 3.4 Decision Fusion . 42 3.4.1 Majority Voting . 44 3.4.2 Sequential Conditional Application . 44 3.5 Mobile Software Framework . 44 4 Results ..................................... 49 4.1 Asymmetry . 49 4.1.1 Data Preprocessing . 49 4.1.2 DermoScreen Asymmetry Analysis . 50 4.1.3 Asymmetry from Border and Area . 51 ix 4.1.4 Asymmetry using Size Functions . 56 4.1.5 Summary of Asymmetry Classifier Performance . 67 4.1.6 Effect of Improved Asymmetry on Melanoma Detection . 68 4.2 Decision Fusion . 71 4.2.1 Decision Fusion Based on Voting Strategies . 72 4.2.2 Decision Fusion Based on Sequential Application . 72 4.3 Mobile Software Framework . 73 4.3.1 Domain Specific Language (DSL) Implementation . 73 4.3.2 Performance Testing . 75 5 Conclusions .................................. 78 5.1 Contributions . 78 5.1.1 Objective 1: Develop an adaptable software framework . 78 5.1.2 Objective 2: Explore methods for assessing asymmetry . 80 5.1.3 Objective 3: Explore decision fusion methods . 81 5.1.4 Additional contributions . 82 5.2 Future work . 83 5.2.1 Software framework . 83 5.2.2 Asymmetry detection . 83 Bibliography ................................... 84 Appendices .................................... 93 A Publications and Patents .......................... 94 x List of Figures 2.1 Dermline II dermatoscope . 7 2.2 The ABCD rule . 9 2.3 7-Point checklist . 10 2.4 Typical analysis architecture for medical image processing . 11 2.5 MCS topologies . 13 2.6 Dermoscopic analysis on iPhone . 20 3.1 Support vector machines . 28 3.2 A simple decision tree . 29 3.3 Receiver operating characteristic curve . 32 3.4 Lesion quadrants based on centroid . 34 3.5 Size functions for a simple curve . 36 3.6 Matching distance between two size functions . 37 3.7 Pixels used in calculating measuring functions . 38 3.8 Tangent values on lesion boundary . 40 3.9 Areas used to determine area irregularity measure . 42 3.10 Conditional application of dermoscopic rules . 45 3.11 Conceptual architecture for proposed imaging framework . 47 4.1 Correlation plot for area and contour features . 53 4.2 Boruta plot for all features using the PHm data set. 54 4.3 ROC curves for Boruta features . 55 xi 4.4 Border distance size functions . 58 4.5 Correlation plot for measuring functions . 59 4.6 Boruta diagram for size function features . 60 4.7 ROC curves for Boruta for size functions . 61 4.8 Correlation plot for combined features . 63 4.9 Boruta diagram for combined features . 64 4.10 ROC curves for LR classifiers and full feature set . 66 4.11 Asymmetry classifier accuracy . 67 4.12 Classification of melanoma using the ABCD rule. 69 4.13 Classification of melanoma using the Menzies rule. 70 4.14 Benign lesion with interal asymmetry . 71 4.15 IPC selection screen . 74 4.16 Task object model . 75 4.17 DSL Example: Creating a mask image . 75 xii List of Tables 2.1 Common melanoma scoring schemes . 8 3.1 Features and their corresponding area descriptors . 43 4.1 PH2 dataset after removal of poor images . 50 4.2 Classification results for Dermoscreen and PH2m dataset . 51 4.3 Features and their abbreviations . 52 4.4 CI and VIF measures for features with CI 30 . 52 ≥ 4.5 Classification results for PH2m dataset . 56 4.6 Classification results for PH2m dataset using size function approach . 62 4.7 Classification results for PH2m dataset using combined feature set . 65 4.8 Results for the LR classifier on the PH20;1 and PH20;2 datasets. 65 4.9 Summary of asymmetry accuracy for LASSO . 67 4.10 Classifier performance on melanoma images . 68 4.11 Summary of dermoscopic rule classifier results . 72 4.12 Results for simple voting strategies . 72 4.13 Fast median filtering performance . 76 xiii Chapter 1 Introduction An estimated two million people are diagnosed with some form of skin cancer annually [60]. These cancers include malignant melanoma, basal cell carcinoma, and squamous cell carcinoma. Melanoma represents only 5% of all new cases, however it accounts for the vast majority of skin cancer deaths globally [49]. In the United States alone, melanoma was estimated to claim 9,730 lives in 2016 [60]. Melanoma in situ, a very early stage of melanoma, can add tens of thousands of more cases, as it is growing at 15% per year [63] . Five-year survival rates for patients whose melanoma is detected early is about 98%. If detection is much later, after the cancer has spread to the internal organs, the rate drops to only 15% [60]. It is readily apparent that early detection is critical to patient survival. A substantial percentage of melanomas are initially detected through self-examination [6][51], using the checklists recommended by the American Academy of Dermatology. It is difficult, however, for even trained physicians to differentiate between melanoma and other pigmented skin lesions solely by visual inspection [48][10]. As a result, physicians routinely order biopsies if the lesion looks suspicious. The inaccuracy of visual inspection methods has motivated the development of automated screening systems that can recognize melanoma at an early stage. If suffi- ciently accurate, these systems have the potential to detect melanoma at a treatable stage, while eliminating at least some of the unnecessary invasive procedures such as scraping or excision.