Auto-Segmentation for Radiation Oncology: State of the Art (Series in Medical Physics and Biomedical Engineering)

"This book provides a comprehensive introduction to current state-of-the-art auto-segmentation approaches used in radiation oncology for auto-delineation of organs-of-risk for thoracic radiation treatment planning. Containing the latest, cutting edge technologies and treatments, it explores deep-learning methods, multi-atlas-based methods, and model-based methods that are currently being developed for clinical radiation oncology applications. Each chapter focuses on a specific aspect of algorithm choices and discusses the impact of the different algorithm modules to the algorithm performance as well as the implementation issues for clinical use (including data curation challenges and auto-contour evaluations). This book is an ideal guide for radiation oncology centers looking to learn more about potential auto-segmentation tools for their clinic in addition to medical physicists commissioning auto-segmentation for clinical use. Features: Up-to-date with the latest technologies in the field Edited by leading authorities in the area, with chapter contributions from subject area specialists All approaches presented in this book are validated using a standard benchmark dataset established by the Thoracic Auto-segmentation Challenge held as an event of the 2017 Annual Meeting of American Association of Physicists in Medicine"-- Provided by publisher Cover Half Title Series Page About the Series Title Page Copyright Page Table of Contents Foreword I Foreword II Editors Contributors Chapter 1 Introduction to Auto-Segmentation in Radiation Oncology 1.1 Introduction 1.2 Evolution of Auto-Segmentation 1.3 Evaluation of Auto-Segmentation 1.4 Benchmark Dataset 1.5 Clinical Implementation Concerns References Part I Multi-Atlas for Auto-Segmentation Chapter 2 Introduction to Multi-Atlas Auto-Segmentation 2.1 Introduction 2.2 Database Construction 2.3 Atlas Selection 2.4 Query Image Registration 2.5 Label Fusion 2.6 Label Post-Processing 2.7 Summary of This Part of the Book References Chapter 3 Evaluation of Atlas Selection: How Close Are We to Optimal Selection? 3.1 Motivation for Atlas Selection 3.2 Methods of Atlas Selection 3.3 Evaluation of Image-Based Atlas Selection 3.3.1 Implementation 3.3.2 Brute-Force Search 3.3.3 Atlas Selection Performance Assessment 3.3.4 Discussion and Implications for Atlas Selection 3.3.5 Limitations 3.4 Impact of Atlas Selection on Clinical Practice 3.5 Summary and Recommendations for Future Research References Chapter 4 Deformable Registration Choices for Multi-Atlas Segmentation 4.1 Introduction 4.1.1 Deformable Registration 4.1.2 B-Spline Registration 4.1.3 Demons Algorithm 4.2 Plastimatch MABS Implementation Details 4.3 Evaluation Metrics 4.4 Experimental steps 4.5 Results 4.6 Summary References Chapter 5 Evaluation of a Multi-Atlas Segmentation System 5.1 Introduction 5.2 Methods 5.2.1 Patient Data 5.2.2 Online Atlas Selection for Multi-Atlas Segmentation 5.2.2.1 First Phase of Atlas Selection 5.2.2.2 Deformable Image Registration 5.2.2.3 Second Phase of Atlas Selection 5.2.2.4 Contour Fusion 5.2.3 Evaluation Metrics 5.2.4 Esophagus Segmentation for Head and Neck Cancer Patients 5.2.5 Validation Using Public Benchmark Dataset 5.3 Results 5.3.1 Atlas Ranking and Selection 5.3.2 Esophagus Segmentation for Head and Neck Cancer Patients 5.3.3 Validation with Public Benchmark Dataset 5.4 Discussion 5.5 Conclusions References Part II Deep Learning for Auto-Segmentation Chapter 6 Introduction to Deep Learning-Based Auto-Contouring for Radiotherapy 6.1 Introduction 6.2 Historical Context 6.2.1 Artificial Neural Networks 6.2.1.1 Convolution Neural Networks 6.2.1.2 Computational Power 6.3 What Makes Deep Learning-Based Contouring So Different to Atlas-Based or Model-Based Approaches 6.3.1 Underlying Assumptions 6.3.2 Use of Data 6.3.3 Degrees of Freedom 6.4 Summary of This Part of the Book References Chapter 7 Deep Learning Architecture Design for Multi-Organ Segmentation 7.1 Introduction 7.2 Deep Learning Architecture in Medical Image Multi-Organ Segmentation 7.2.1 Auto-Encoder Methods 7.2.1.1 Auto-Encoder and Its Variants 7.2.1.2 Overview of Works 7.2.1.3 Discussion 7.2.2 CNN Methods 7.2.2.1 Network Designs 7.2.2.2 Overview of Works 7.2.2.3 Discussion 7.2.3 FCN Methods 7.2.3.1 Network Designs 7.2.3.2 Overview of Works 7.2.3.3 Discussion 7.2.4 GAN Methods 7.2.4.1 Network Designs 7.2.4.2 Overview of Works 7.2.4.3 Discussion 7.2.5 R-CNN Methods 7.2.5.1 Network Designs 7.2.5.2 Overview of Works 7.2.5.3 Discussion 7.2.6 Hybrid Methods 7.2.6.1 Network Designs 7.2.6.2 Overview of Works 7.2.6.3 Discussion 7.3 Benchmark 7.4 Conclusion Acknowledgments References Chapter 8 Comparison of 2D and 3D U-Nets for Organ Segmentation 8.1 Introduction 8.2 Structures of 2D and 3D U-Nets 8.2.1 2D U-Net 8.2.2 3D U-Net 8.3 Experimental Results 8.3.1 Datasets 8.3.2 Evaluation Metrics 8.3.3 Implementation Details 8.3.4 Results 8.3.5 Discussions 8.4 Summary References Chapter 9 Organ-Specific Segmentation Versus Multi-Class Segmentation Using U-Net 9.1 Introduction 9.2 Materials and Methods 9.2.1 Datasets 9.2.2 Network Structure 9.2.3 Pre-Processing and Downsampling 9.2.4 Quantitative Evaluation Metrics 9.2.5 Implementation and Comparison Experiments 9.3 Results 9.4 Discussion 9.5 Conclusions Acknowledgments References Chapter 10 Effect of Loss Functions in Deep Learning-Based Segmentation 10.1 Introduction 10.2 Admissibility of a Loss Function 10.3 Presenting the Problem 10.4 Common Loss Functions 10.4.1 Mean Squared Error 10.4.1.1 Cross Entropy 10.4.1.2 Binary Cross Entropy 10.4.1.3 Categorical Cross Entropy 10.4.2 Dice Loss 10.4.3 Hausdorff Distance Loss 10.5 Dealing with Class Imbalance 10.5.1 Weighted Cross Entropy 10.5.2 Generalized Dice Loss 10.5.3 No-Background Dice Loss 10.5.4 Focal Loss 10.5.5 Sensitivity Specificity Loss 10.5.6 Tversky Loss 10.6 Compound Loss Functions 10.6.1 Dice + Cross Entropy 10.6.2 Dice + Focal Loss 10.6.3 Non-Linear Combinations 10.7 Dealing with Imperfect Data 10.8 Evaluating a Loss Function References Chapter 11 Data Augmentation for Training Deep Neural Networks 11.1 Overview 11.2 Introduction and Literature Review 11.3 Geometric Transformations 11.4 Intensity Transformation 11.5 Artificial Data Generation 11.6 Applications of Data Augmentation 11.6.1 Datasets and Image Preprocessing 11.6.2 Training, Validation, and Testing for Organ Segmentation 11.7 Evaluation Criteria 11.8 Results 11.9 Discussion 11.10 Summary Acknowledgments References Chapter 12 Identifying Possible Scenarios Where a Deep Learning Auto-Segmentation Model Could Fail 12.1 Background 12.1.1 Site-Specific Models 12.1.2 Limitations of Training Data 12.2 Deep Learning Architecture 12.2.1 Two-Stage U-Net Model 12.2.1.1 Image Preprocessing 12.2.1.2 Stage 1: Localization through Coarse Segmentations 12.2.1.3 Stage 2: OAR Segmentation through Fine-Detail Segmentation 12.2.2 Test-Time Cluster Cropping Technique 12.3 Quantitative and Qualitative Review of Auto-Segmentations 12.3.1 Performance on Challenge Test Set 12.3.2 Different Anatomical Sites 12.3.2.1 Head and Neck Scans 12.3.2.2 Abdominal Scans 12.3.2.3 Breast Cancer Simulation Scans 12.3.3 Different Clinical Presentations 12.3.3.1 Atelectasis and Pleural Effusion 12.3.3.2 Presence of Motion Management Devices 12.3.3.3 Use of Contrast and the Presence of Implanted Devices 12.3.3.4 Adapting to the Unseen 12.4 Discussion and Conclusions Acknowledgments References Part III Clinical Implementation Concerns Chapter 13 Clinical Commissioning Guidelines 13.1 Introduction 13.2 Stages in Clinical Commissioning 13.3 Need for Robust and Clinically Useful Metrics 13.4 Need for Curated Datasets for Clinical Commissioning 13.5 Auto-Segmentation Clinical Validation Studies – Current State-of-the-Art 13.6 Data Curation Guidelines for Radiation Oncology 13.7 Evaluation Metrics Guidelines for Clinical Commissioning 13.8 Commissioning and Safe Use in the Clinic 13.8.1 Training and Validation Phase 13.8.2 Testing and Verification Phase 13.9 Summary References Chapter 14 Data Curation Challenges for Artificial Intelligence 14.1 Introduction 14.2 The Complexity of Medical Imaging 14.3 The Challenge of Generalizability and Data Heterogeneity 14.4 Data Selection 14.5 The Need for Large Quantities of Data 14.6 Barriers to Sharing Patient Data and Distributed Deep Learning 14.7 Data Quality Issues 14.8 Data Annotations 14.9 Data Curation via Competitions 14.10 Bias and Curation of Fair Data 14.11 Overview of Data Curation Process 14.12 Conclusion References Chapter 15 On the Evaluation of Auto-Contouring in Radiotherapy 15.1 Introduction 15.2 Quantitative Evaluation 15.2.1 Strengths and Limitations 15.2.2 Implementation 15.2.3 Classification Accuracy 15.2.3.1 Implementation 15.2.3.2 Advantages and Limitations 15.2.4 Dice Similarity Coefficient 15.2.4.1 Implementation 15.2.4.2 Advantages and Limitations 15.2.5 Distance Measures 15.2.5.1 Hausdorff Distance 15.2.5.2 95% Hausdorff Distance 15.2.5.3 Average Distance 15.2.5.4 Implementation 15.2.5.5 Advantages and Limitations 15.2.6 Geometric Properties 15.2.6.1 Centroid Location Comparison 15.2.6.2 Volume Comparison 15.2.6.3 Implementation 15.2.6.4 Advantages and Limitations 15.2.7 Measures of Estimated Editing 12.2.7.1 Implementation 15.2.7.2 Advantages and Limitations 15.2.8 Handling Inter-Observer Variation in Quantitative Assessment 15.2.9 Summary of Quantitative Evaluation 15.2.10 Example Implementation 15.3 Subjective Evaluation 15.3.1 The Acceptance of Contours 15.3.2 The Source of Contouring 15.3.3 The Preference for Contouring 15.3.4 Challenges of Subjective Assessment 15.3.5 Summary of Subjective Evaluation 15.4 Assessment Based on Intended Clinical Use 15.4.1 Evaluation of Time Saving 15.4.1.1 Challenges for Study Design 15.4.2 Impact of Auto-Contouring on Planning 15.4.2.1 Challenges for Study Design 15.4.3 Summary of Clinical Impact Evaluation 15.5 Discussion and Recommendations References Index