Auto-Segmentation for Radiation Oncology: State of the Art - Auto-Segmentation for Radiation Oncology: State of the Art -

Introduction Evolution of Auto-Segmentation Evaluation of Auto-Segmentation Benchmark Dataset Clinical Implementation Concerns References I Multi-Atlas for Auto-SegmentationIntroduction to Multi-Atlas Auto-Segmentation Introduction Database Construction Atlas Selection Query Image Registration Label Fusion Label Post-Processing Summary of This Part of the Book References Evaluation of Atlas Selection: How Close Are We to Optimal Selection?Motivation for Atlas Selection Methods of Atlas Selection Evaluation of Image-Based Atlas Selection Implementation Brute-Force Search Atlas Selection Performance Assessment Discussion and Implications for Atlas Selection Limitations Impact of Atlas Selection on Clinical Practice Summary and Recommendations for Future Research References Deformable Registration Choices for Multi-Atlas Segmentation Introduction Deformable Registration B-Spline Registration Demons Algorithm Plastimatch MABS Implementation Details Evaluation Metrics Experimental steps Results Summary References Evaluation of a Multi-Atlas Segmentation System Introduction Methods Patient Data Online Atlas Selection for Multi-Atlas Segmentation First Phase of Atlas Selection Deformable Image Registration Second Phase of Atlas Selection Contour Fusion Evaluation Metrics Esophagus Segmentation for Head and Neck Cancer Patients Validation Using Public Benchmark Dataset Results Atlas Ranking and Selection Esophagus Segmentation for Head and Neck Cancer Patients Validation with Public Benchmark Dataset Discussion Conclusions References II Deep Learning for Auto-SegmentationIntroduction to Deep Learning-Based Auto-Contouring for Radiotherapy Introduction Historical Context Artificial Neural Networks Convolution Neural Networks Computational Power What Makes Deep Learning-Based Contouring So Different to Atlas-Based or Model-Based Approaches Underlying Assumptions Use of Data Degrees of Freedom Summary of This Part of the Book References Deep Learning Architecture Design for Multi-Organ Segmentation Introduction Deep Learning Architecture in Medical Image Multi-Organ Segmentation Auto-Encoder Methods Auto-Encoder and Its Variants Overview of Works Discussion CNN Methods Network Designs Overview of Works Discussion FCN Methods Network Designs Overview of Works Discussion GAN Methods Network Designs Overview of Works Discussion R-CNN Methods Network Designs Overview of Works Discussion Hybrid Methods Network Designs Overview of Works Discussion Benchmark Conclusion Acknowledgments References Comparison of 2D and 3D U-Nets for Organ Segmentation Introduction Structures of 2D and 3D U-Nets D U-Net D U-Net Experimental Results Datasets Evaluation Metrics Implementation Details Results Discussions Summary References Organ-Specific Segmentation Versus Multi-Class Segmentation Using U-Net Introduction Materials and Methods Datasets Network Structure Pre-Processing and Downsampling Quantitative Evaluation Metrics Implementation and Comparison Experiments Results Discussion Conclusions Acknowledgments References Effect of Loss Functions in Deep Learning-Based Segmentation Introduction Admissibility of a Loss Function Presenting the Problem Common Loss Functions Mean Squared Error Cross Entropy Binary Cross Entropy Categorical Cross Entropy Dice Loss Hausdorff Distance Loss Dealing with Class Imbalance Weighted Cross Entropy Generalized Dice Loss No-Background Dice Loss Focal Loss Sensitivity Specificity Loss Tversky Loss Compound Loss Functions Dice + Cross Entropy Dice + Focal Loss Non-Linear Combinations Dealing with Imperfect Data Evaluating a Loss Function References Data Augmentation for Training Deep Neural Networks Overview Introduction and Literature Review Geometric Transformations Intensity Transformation Artificial Data Generation Applications of Data Augmentation Datasets and Image Preprocessing Training, Validation, and Testing for Organ Segmentation Evaluation Criteria Results Discussion Summary Acknowledgments References Identifying Possible Scenarios Where a Deep Learning Auto-Segmentation Model Could Fail Background Site-Specific Models Limitations of Training Data Deep Learning Architecture Two-Stage U-Net Model Image Preprocessing Stage 1: Localization through Coarse Segmentations Stage 2: OAR Segmentation through Fine-Detail Segmentation Test-Time Cluster Cropping Technique Quantitative and Qualitative Review of Auto-Segmentations Performance on Challenge Test Set Different Anatomical Sites Head and Neck Scans Abdominal Scans Breast Cancer Simulation Scans Different Clinical Presentations Atelectasis and Pleural Effusion Presence of Motion Management Devices Use of Contrast and the Presence of Implanted Devices Adapting to the Unseen Discussion and Conclusions Acknowledgments References III Clinical Implementation ConcernsClinical Commissioning Guidelines Introduction Stages in Clinical Commissioning Need for Robust and Clinically Useful Metrics Need for Curated Datasets for Clinical Commissioning Auto-Segmentation Clinical Validation Studies – Current State-of-the-Art Data Curation Guidelines for Radiation Oncology Evaluation Metrics Guidelines for Clinical Commissioning Commissioning and Safe Use in the Clinic Training and Validation Phase Testing and Verification Phase Summary References Data Curation Challenges for Artificial Intelligence Introduction The Complexity of Medical Imaging The Challenge of Generalizability and Data Heterogeneity Data Selection The Need for Large Quantities of Data Barriers to Sharing Patient Data and Distributed Deep Learning Data Quality Issues Data Annotations Data Curation via Competitions Bias and Curation of Fair Data Overview of Data Curation Process Conclusion References On the Evaluation of Auto-Contouring in Radiotherapy Introduction Quantitative Evaluation Strengths and Limitations Implementation Classification Accuracy Implementation Advantages and Limitations Dice Similarity Coefficient Implementation Advantages and Limitations Distance Measures Hausdorff Distance % Hausdorff Distance Average Distance Implementation Advantages and Limitations Geometric Properties Centroid Location Comparison Volume Comparison Implementation Advantages and Limitations Measures of Estimated Editing Implementation Advantages and Limitations Handling Inter-Observer Variation in Quantitative Assessment Summary of Quantitative Evaluation Example Implementation Subjective Evaluation The Acceptance of Contours The Source of Contouring The Preference for Contouring Challenges of Subjective Assessment Summary of Subjective Evaluation Assessment Based on Intended Clinical Use Evaluation of Time Saving Challenges for Study Design Impact of Auto-Contouring on Planning Challenges for Study Design Summary of Clinical Impact Evaluation Discussion and Recommendations References