Deep Network Design for Medical Image Computing : Principles and Applications

Deep Network Design for Medical Image Computing: Principles and Applications covers a range of MIC tasks and discusses design principles of these tasks for deep learning approaches in medicine. These include skin disease classification, vertebrae identification and localization, cardiac ultrasound image segmentation, 2D/3D medical image registration for intervention, metal artifact reduction, sparse-view artifact reduction, etc. For each topic, the book provides a deep learning-based solution that takes into account the medical or biological aspect of the problem and how the solution addresses a variety of important questions surrounding architecture, the design of deep learning techniques, when to introduce adversarial learning, and more. This book will help graduate students and researchers develop a better understanding of the deep learning design principles for MIC and to apply them to their medical problems. Explains design principles of deep learning techniques for MIC Contains cutting-edge deep learning research on MIC Covers a broad range of MIC tasks, including the classification, detection, segmentation, registration, reconstruction and synthesis of medical images Front Cover Deep Network Design for Medical Image Computing Copyright Contents List of figures Acknowledgments 1 Introduction 1.1 Medical image computing 1.1.1 Medical image reconstruction 1.1.2 Medical image analysis 1.1.3 Medical image computing as functional approximation 1.2 Deep learning design principles 1.2.1 Computer vision techniques for medical image computing 1.2.2 Machine learning techniques for medical image computing 1.2.3 Medical domain knowledge 1.3 Chapter organization References 2 Deep learning basics 2.1 Convolutional neural networks 2.1.1 3D convolutional neural networks 2.2 Recurrent neural networks 2.2.1 Long short-term memory 2.2.2 Bidirectional RNN 2.3 Deep image-to-image networks 2.3.1 Retaining spatial resolutions 2.3.2 Fully convolutional networks 2.3.3 Encoder–decoder networks 2.4 Deep generative networks 2.4.1 Basic models References Part 1 Deep network design for medical image analysis and selected applications 3 Classification: lesion and disease recognition 3.1 Design principles 3.1.1 Choice of deep neural networks 3.1.2 Choice of classification tasks and objectives 3.1.3 Transfer learning 3.1.4 Multitask learning 3.2 Case study: skin disease classification versus skin lesion characterization 3.2.1 Background 3.2.2 Dataset 3.2.3 Methodology 3.2.4 Experiments 3.2.5 Discussion 3.3 Case study: skin lesion classification with multitask learning 3.3.1 Background 3.3.2 Dataset 3.3.3 Methodology 3.3.4 Experiments 3.3.5 Discussion 3.4 Summary References 4 Detection: vertebrae localization and identification 4.1 Design principles 4.1.1 Choice of deep neural networks 4.1.2 Choice of detection tasks and objectives 4.2 Case study: vertebrae localization and identification 4.2.1 Background 4.2.2 Methodology 4.2.3 Experiments 4.2.4 Discussion 4.3 Summary References 5 Segmentation: intracardiac echocardiography contouring 5.1 Design principles 5.1.1 Choice of deep neural networks 5.1.2 Choice of segmentation tasks and objectives 5.1.3 Image restoration for segmentation 5.2 Case study: intracardiac echocardiography contouring 5.2.1 Methodology 5.2.2 Experiments 5.2.3 Discussion 5.3 Summary References 6 Registration: 2D/3D rigid registration 6.1 Design principles 6.1.1 Deep similarity based registration 6.1.1.1 Problem definition and choice of objective functions 6.1.1.2 Deep learning models for similarity metric learning 6.1.2 Reinforcement learning based registration 6.1.2.1 Problem definition and choice of objective functions 6.1.2.2 Deep learning models for reinforcement learning based registration 6.1.3 Supervised transformation estimation 6.1.3.1 Problem definition and choice of objective functions 6.1.3.2 Deep learning models for supervised transformation estimation 6.1.4 Unsupervised transformation estimation 6.1.4.1 Problem definition and choice of objective functions 6.1.4.2 Deep learning models for unsupervised transformation estimation 6.2 Case study: 2D/3D medical image registration 6.2.1 Problem formulation 6.2.2 Methodology 6.2.3 Experiments 6.2.4 Limitations 6.2.5 Discussion 6.3 Summary References Part 2 Deep network design for medical image reconstruction, synthesis, and selected applications 7 Reconstruction: supervised artifact reduction 7.1 Design principles 7.1.1 Image domain approaches 7.1.1.1 Problem definition and choice of objective functions 7.1.1.2 Deep learning models for image domain reconstruction 7.1.2 Sensor domain approaches 7.1.2.1 Problem definition and choice of objective functions 7.1.2.2 Deep learning models for sensor domain reconstruction 7.1.3 Dual-domain approaches 7.1.3.1 Problem definition and choice of objective functions 7.1.3.2 Deep learning models for dual-domain reconstruction 7.2 Case study: sparse-view artifact reduction 7.2.1 Background 7.2.2 Methodology 7.2.2.1 Network structure 7.2.2.2 Focus map 7.2.3 Experiments 7.2.3.1 Dataset and models 7.2.3.2 Results 7.2.4 Discussion 7.3 Case study: metal artifact reduction 7.3.1 Background 7.3.1.1 Inpainting-based methods 7.3.1.2 MAR by iterative reconstruction 7.3.2 Methodology 7.3.2.1 Sinogram enhancement network 7.3.2.2 Radon inversion layer 7.3.2.3 Image enhancement network 7.3.3 Experiments 7.3.3.1 Ablation study 7.3.3.2 Comparison with state-of-the-art methods 7.3.3.3 Running time comparisons 7.3.4 Discussion 7.4 Summary References 8 Reconstruction: unsupervised artifact reduction 8.1 Design principles 8.1.1 Unpaired learning approaches 8.1.1.1 Problem definition and choice of objective functions 8.1.1.2 Deep learning models for unpaired learning of medical image reconstruction 8.1.2 Self-supervised learning approaches 8.1.2.1 Problem definition and choice of objective functions 8.1.2.2 Deep learning models for self-supervised learning of medical image reconstruction 8.2 Case study: metal artifact reduction 8.2.1 Background 8.2.2 Methodology 8.2.2.1 Encoders and decoders 8.2.2.2 Learning 8.2.2.3 Network architectures 8.2.3 Experiments 8.2.3.1 Baselines 8.2.3.2 Datasets 8.2.3.3 Training and testing 8.2.3.4 Performance on synthesized data 8.2.3.5 Performance on clinical data 8.2.3.6 Ablation study 8.2.3.7 Artifact synthesis 8.2.4 Discussion 8.3 Summary References 9 Synthesis: novel radiography view synthesis 9.1 Design principles 9.1.1 Unconditional synthesis 9.1.1.1 Problem definition and choice of objective functions 9.1.1.2 Deep learning models for unconditional medical image synthesis 9.1.2 Homogeneous domain synthesis 9.1.2.1 Problem definition and choice of objective functions 9.1.2.2 Deep learning models for homogeneous domain synthesis 9.1.3 Heterogeneous domain synthesis 9.1.3.1 Deep learning models for heterogeneous domain synthesis 9.2 Case study: novel radiography view synthesis 9.2.1 Background 9.2.1.1 View synthesis from a single image 9.2.1.2 Radiograph simulation and transformation to CT 9.2.2 Methodology 9.2.2.1 CT2Xray 9.2.2.2 XraySyn 9.2.3 Experiments 9.2.3.1 Implementation details 9.2.3.2 Dataset 9.2.3.3 Evaluation metrics 9.2.3.4 Ablation study 9.2.3.5 Bone suppression 9.2.4 Discussion 9.3 Summary References 10 Challenges and future directions 10.1 Challenges and open issues 10.1.1 Effectiveness in clinical workflows 10.1.2 Responsible AI for healthcare 10.2 Trends and future directions References Index Back Cover