Machine and Deep Learning in Oncology, Medical Physics and Radiology

Foreword to the First Edition
Preface to the First Edition
Preface to the Second Edition
Contents
Part I: Introduction to Machine and Deep Learning Principles
1: What Are Machine and Deep Learning?
1.1 Overview
1.2 Background
1.3 Machine Learning Definition
1.4 Deep Learning Definition
1.5 Learning from Data
1.6 Overview of Machine and Deep Learning Approaches
1.7 Quantifying the Data and Learning Objectives
1.8 Application in Biomedicine
1.9 Applications in Radiology and Oncology
1.10 Ethical Challenges in the Application of Machine Learning
1.11 Steps to Machine Learning Heaven
1.12 Conclusions
References
2: Computational Learning Theory
2.1 Introduction
2.2 Computational Modeling Versus Statistics
2.3 Learning Capacity
2.4 PAC Learning
2.5 VC Dimension
2.6 Learning with Deep Learning
2.7 Model Complexity Analysis in Practice
2.7.1 Model Order Based on Information Theory
2.7.2 Model Order Based on Resampling Methods
2.8 Conclusions
References
3: Conventional Machine Learning Methods
3.1 Introduction
3.2 Unsupervised Learning
3.2.1 Linear Principal Component Analysis
3.2.2 Kernel Principal Component Analysis
3.2.3 Factor Analysis (FA)
3.2.4 Clustering
3.3 Supervised Learning
3.3.1 Logistic Regression
3.3.2 Feed-Forward Neural Networks (FFNN)
3.3.3 General Regression Neural Networks (GRNN)
3.3.4 Kernel-Based Methods
3.3.5 Decision Trees and Random Forests
3.3.6 Bayesian Network
3.3.7 Naive Bayes
3.4 Reinforcement Learning
3.4.1 Reinforcement Learning for Adaptive Liver Cancer Treatment
References
4: Overview of Deep Machine Learning Methods
4.1 Introduction
4.2 The Vanilla Neural Network
4.2.1 Training a Neural Network
4.2.2 Hyperparameters Associated with Training
4.2.3 What Makes a Neural Network Deep?
4.2.4 Example: Neural Network for Binary Classification
4.3 Autoencoders
4.4 Convolutional Neural Networks (CNNs)
4.4.1 Convolutions
4.4.2 Pooling
4.5 Recurrent Neural Networks
4.5.1 Long Short-Term Memory (LSTM)
4.5.2 Gated Recurrent Units (GRUs)
4.6 Generative Adversarial Networks (GANs)
4.6.1 Vanilla GANs
4.6.2 Common GAN Variants: DCGAN, WGAN
4.7 Deep Reinforcement Learning (DRL)
4.8 Current Challenges and Future Directions
4.9 Conclusion
References
5: Quantum Computing for Machine Learning
5.1 Introduction
5.2 Postulates of Quantum Mechanics
5.3 Quantum Hardware
5.3.1 Quantum Annealers
5.3.2 Universal Quantum Computers
5.4 Common Quantum Computing Algorithms
5.4.1 Grover’s Algorithm
5.4.2 Quantum Phase Estimation
5.4.3 Shor’s Algorithm
5.4.4 Quantum Machine Learning
5.4.4.1 Quantum Support Vector Machines
5.4.4.2 Quantum Principal Component Analysis
5.4.4.3 Quantum Bayesian Network
5.4.4.4 Quantum Neural Network and Deep Learning
5.4.4.5 Quantum Reinforcement Learning
5.5 Application of Quantum Computing in Medical Physics
5.5.1 Optimization and Planning
5.5.2 Outcome Modeling/Decision Making
5.6 Conclusion
References
6: Performance Evaluation
6.1 Standard Evaluation Methods for Machine Learning Systems
6.1.1 Choosing an Appropriate Performance Measure
6.1.1.1 Common Metrics Used in all Machine Learning Applications
6.1.1.2 Metrics Used Specifically in Medical Machine Learning Applications
6.1.1.3 Common Metrics Used in Computer Imaging Applications
6.1.2 Choosing an Appropriate Sampling Method
6.1.3 Choosing an Appropriate Statistical Testing Strategy
6.1.3.1 In the Context of a Single Classifier
6.1.3.2 In the Context of Several Classifiers
6.2 Standard Practice in Medical Imaging and Oncology
6.2.1 Review of the Current Practice in Medical Imaging and Oncology
6.2.2 Areas where Improvements Could Be Made
6.2.3 Lessons from the Past
References
7: Software Tools for Machine and Deep Learning
7.1 Introduction
7.2 Python-Based Machine Learning Library
7.2.1 Pip and Conda
7.2.2 NumPy and SciPy
7.2.3 Dedicated Machine Learning Libraries
7.2.3.1 Scikit-Learn
7.2.3.2 Shogun
7.2.3.3 mlpy
7.2.3.4 PyMVPA
7.2.3.5 MDP
7.2.3.6 PyBrain
7.2.4 Deep Learning
7.2.4.1 Theano
7.2.4.2 Chainer
7.2.4.3 TensorFlow
7.2.4.4 PyTorch
7.2.4.5 Caffe
7.2.4.6 MXNet
7.2.5 Examples
7.2.6 Benchmark
7.3 Weka
7.4 R
7.5 Matlab
7.6 Cloud-Based Platforms
7.6.1 AWS Deep Learning AMIs and SageMaker
7.6.2 Google Colab
7.6.3 Azure Machine Learning Studio
7.6.4 IBM Watson Machine Learning Studio
7.7 Conclusions
References
8: Privacy-Preserving Federated Data Analysis: Data Sharing, Protection, and Bioethics in Healthcare
8.1 Introduction
8.1.1 Data Landscape
8.1.1.1 Structured Data and Unstructured Data
8.1.1.2 Horizontally Partitioned Data and Vertically Partitioned Data
8.2 Prerequisites
8.2.1 Data Extraction
8.2.1.1 ETL Tooling and Data Warehousing
8.2.1.2 Image Biomarker Extraction
8.2.2 Data Representation and FAIR Data Principles
8.2.2.1 Relational Databases and Ontologies
8.2.2.2 Semantic Web, RDF, and Linked Data
Resource Description Framework
Unique Resource Identifiers and Linked Data
Querying Using SPARQL
8.2.2.3 HL7 FHIR and REST-APIs
8.2.3 Network Infrastructure
8.2.3.1 Institutional Infrastructure
Traditional ETL and DWH
FAIR Data Store
Traditional ETL and DWH with a FAIR Store
Traditional ETL and DWH with a Virtual FAIR Store
Virtual FAIR Store per Institute
Virtual FAIR Store per Source and Institute
8.2.3.2 Machine Learning Infrastructure
Centralized Machine Learning Infrastructure
Distributed Machine Learning Infrastructure: The Personal Health Train
8.2.4 Centralized and Distributed Machine Learning Algorithms
8.2.4.1 Centralized Machine Learning
8.2.4.2 Distributed Machine Learning
Horizontal Distributed (Federated) Machine Learning
Vertical Distributed (Federated) Learning
8.2.5 Bioethics and Data Protection
8.2.5.1 Bioethics and Data Protection: Individuals
8.2.5.2 Bioethics and Data Protection: Data Entity
Pseudonymization
Data Obfuscation
Data Perturbation
8.2.5.3 Bioethics and Data Protection: Society
8.2.6 Applications and Initiatives
8.2.6.1 Datashield
8.2.6.2 I2B2
8.2.6.3 VATE
8.2.6.4 PCORnet
8.2.6.5 FAIRHealth
8.2.6.6 Personal Health Train Initiatives
EuroCAT
20 K Challenge
8.2.7 Summary
References
Part II: Machine Learning for Medical Image Analysis in Radiology and Oncology
9: Computerized Detection of Lesions in Diagnostic Images with Early Deep Learning Models
9.1 Introduction
9.2 Overview of Architecture of a CADe Scheme
9.3 Machine Learning (ML) in CADe
9.3.1 Feature-Based (Segmented-Object-Based) ML (Classifiers)
9.3.2 Early Deep Learning Models
9.3.2.1 Overview
9.3.2.2 Difference Between Deep Learning and Feature-Based ML (Classifiers)
9.3.2.3 Early Deep Learning Model: Massive-Training Artificial Neural Network (MTANN)
9.4 CADe in Thoracic Imaging
9.4.1 Thoracic Imaging for Lung Cancer Detection
9.4.2 CADe of Lung Nodules in Thoracic CT
9.4.2.1 Overview
9.4.2.2 Illustration of a CADe Scheme
9.4.3 CADe of Lung Nodules in CXR
9.5 CADe in Colonic Imaging
9.5.1 Colonic Imaging for Colorectal Cancer Detection
9.5.2 Overview of CADe of Polyps in CTC
9.6 Summary
References
10: Classification of Malignant and Benign Tumors
10.1 Introduction
10.2 Overview of Classification Framework
10.2.1 Perception Modeling
10.2.2 Feature Extraction for Tumor Quantification
10.2.3 Design of Decision Function Using Machine Learning
10.2.4 Deep Learning Methods
10.2.5 CADx Classifier Training and Performance Evaluation
10.3 Application Examples in Mammography
10.3.1 Mammography
10.3.2 Detection of Clustered Microcalcifications in Mammograms
10.3.3 Computer-Aided Diagnosis (CADx) of Microcalcification Lesions in Mammograms
10.3.4 Adaptive CADx Boosted with Content-Based Image Retrieval (CBIR)
10.4 MDS as a Visualization Tool of Example Lesions
10.4.1 Multidimensional Scaling (MDS) Technique
10.4.2 Exploring Similar MC Lesions with MDS
10.5 Issues and Recommendations
10.6 Conclusions
References
11: Auto-contouring for Image-Guidance and Treatment Planning
11.1 Introduction
11.2 Traditional Auto-Segmentation Techniques
11.2.1 First-Generation Auto-Segmentation Techniques
11.2.2 Second-Generation Auto-Segmentation Techniques
11.2.3 Third-Generation Auto-Segmentation Techniques
11.3 Deep Learning-Based Auto-Segmentation
11.3.1 Convolutional Neural Networks and Fully Convolutional Networks
11.3.2 Popular Deep Learning Auto-Segmentation Architectures
11.4 Image Segmentation Packages and Publicly Available Datasets
11.4.1 Open-Source Image Segmentation Packages
11.4.2 Publicly Available Datasets
11.4.3 Commercial Systems
11.5 Auto-Segmentation Software Commissioning and Quality Assurance
11.5.1 Auto-Segmentation Evaluation
11.5.2 Patient-Specific Evaluations
11.5.3 Commissioning and QA
11.5.4 Current Limitations to Auto-Segmentation Algorithm Development and Implementation
11.6 Overview of State-of-the-Art Results in Medical Image Auto-Segmentation
11.6.1 Normal Tissues
11.6.1.1 Craniospinal
11.6.1.2 Head and Neck
11.6.1.3 Thoracic
11.6.1.4 Pelvis and Abdomen
11.6.2 Tumors and Clinical Target Volumes
11.6.2.1 Tumors
11.6.2.2 Clinical Target Volumes
11.7 Conclusion
References
Part III: Machine Learning for Radiation Oncology Workflow
12: Machine Learning Applications in Quality Assurance of Radiation Delivery
12.1 Introduction
12.2 Overview of the Use of Machine Learning in Quality Assurance and Treatment Delivery
12.2.1 Automated Chart Review
12.2.2 Machine Learning Applied to Delivery Systems
12.2.3 Machine Learning Applied to IMRT QA
12.3 Future Directions
References
13: Knowledge-Based Treatment Planning
13.1 Introduction
13.2 Anatomical Feature-Based KBP Model
13.2.1 Distance to Target Histogram
13.2.2 Model Training and Validation
13.3 A Robust Ensemble Model with Outlier Filtering Mechanism
13.3.1 An Ensemble KBP Model
13.3.2 Outlier Filtering
13.3.2.1 Anatomical Outliers and Dosimetric Outliers
13.3.2.2 Prediction Performance Measure
13.3.2.3 Model-Based Case Filtering Method
13.3.3 Retrospective Validation
13.4 A KBP Model for Multiple-PTV Plans
13.4.1 Generalized Distance to Target Histogram
13.4.2 Modeling with a gDTH-Based Similarity Metric
13.4.3 Data Augmentation
13.4.4 Training and Validation
13.5 Head and Neck Trade-off KBP Model
13.5.1 Plan Trade-off Modeling
13.5.2 Trade-off Simulation and Validation
13.6 A Complete Workflow for KBP Planning of Whole Breast Radiation Therapy
13.6.1 Digitally Reconstructed Radiograph (DRR)-Based Energy Selection
13.6.2 Anatomy-Driven Fluence Estimation
13.6.3 Patient-Specific Fluence Fine-Tuning
13.6.4 Planning Validation
13.6.4.1 Data Selection
13.6.4.2 Model Training and Validation
13.6.4.3 Plan Quality Comparison
13.6.4.4 Plan Efficiency
13.7 Beam Bouquet Knowledge Model for Lung IMRT Planning
13.7.1 Dissimilarity Metric between Two Beam Bouquets
13.7.2 Establishing the Standardized Beam Bouquets
13.7.3 Validation with Clinical Cases
13.8 Summary
References
14: Intelligent Respiratory Motion Management for Radiation Therapy Treatment
14.1 The Problem of Respiratory Movement During Radiotherapy
14.2 Dynamic Compensation Strategies during Delivery
14.3 Using an Artificial Neural Network (ANN) to Model and Predict Breathing Motion
14.4 Basic Neural Network Architecture for Correlation and Prediction
14.4.1 The Single Neuron, or Linear Filter
14.4.2 The Basic Feed-Forward Artificial Neural Network for Prediction
14.4.3 Training the Feed-Forward Network
14.4.4 The Recurrent Network
14.5 Performance of Basic Neural Networks to Predict Tumor Motion
14.5.1 Breathing Prediction Examples for a Simple Feed-Forward Network
14.6 Advanced Neural Network Architectures
14.6.1 Quadratic Neural Unit
14.6.2 Using a Kalman filter to Predict/Correct as Part of the Training Loop
14.6.3 A Network with Multiple Breathing Signal Inputs
14.6.4 Deep Learning Neural Networks for Prediction
14.7 Support Vector Regression (SVR) as an Alternative to Neural Networks for Breathing Prediction
14.8 Probabilistic Neural Networks
14.9 Summary
References
Part IV: Machine Learning for Outcomes Modeling and Decision Support
15: Prediction of Oncology Treatment Outcomes
15.1 Introduction
15.2 Outcome Modeling in Radiotherapy
15.3 Data Resources
15.3.1 Clinical Data
15.3.2 Dosimetric Data
15.3.3 Radiomics (Imaging Features)
15.3.4 Biological Markers
15.4 Database Technologies for Machine Learning in Oncology
15.5 Pan- Vs. P-OMICs
15.5.1 Spurious Relationship
15.5.2 Echo Chamber Effect
15.5.3 Yule–Simpson Paradox
15.5.4 Ghost Analytics
15.6 Modeling Methods
15.6.1 Bottom-up Approaches for Modeling Oncology Response
15.6.2 Top-Down Approaches for Modeling Oncology Response
15.6.2.1 Logistic Regression
A Logistic Outcome Modeling Example
15.6.2.2 Machine Learning Methods
A Machine Learning Outcome Modeling Example
15.7 Software Tools for Outcome Modeling
15.8 Discussion
15.9 Future Research Directions
15.10 Conclusion
References
16: Radiomics and Radiogenomics
16.1 Introduction
16.2 Technical Basis of Radiomics
16.3 Key Findings and Clinical Applications
16.4 Emerging Paradigms: Deep Learning
16.5 Radiogenomics: Integrating Imaging with Genomics
16.6 Current Challenges and Potential Solutions
16.6.1 Standardization and Quantitative Imaging
16.6.2 Reproducibility and Need for Prospective Validation
16.6.3 Data and Software Sharing
16.7 Conclusion and Future Outlook
References
17: Modelling of Radiotherapy Response (TCP/NTCP)
17.1 Introduction
17.1.1 General Considerations
17.2 Tumour Control Probability
17.3 Machine Learning for TCP Modelling
17.4 Example 1: Dosimetric and Clinical Variables
17.4.1 Data Set
17.4.2 Data Exploration
17.4.3 Logistic Regression Modelling Example
17.4.4 Kernel-Based Modelling Example
17.4.5 Comparison with Other Known Models
17.5 Use of Imaging Features
17.6 Use of Biological Markers
17.7 NTCP Modelling
17.7.1 NTCP Models
17.7.2 Dosimetric Data Reduction-Summary Measure
17.8 Machine Learning Approaches to NTCP Modelling
17.8.1 Multivariable Logistic Regression
17.8.2 Feature Selection
17.9 Classical Machine Learning Approaches
17.9.1 Artificial Neural Networks
17.9.2 Support Vector Machines (SVM)
17.9.3 Unsupervised Learning SOM
17.9.4 Bayesian Networks
17.9.5 Decision Trees
17.9.6 Random Forests
17.9.7 Hybrid Models and Comparative Studies
17.10 Deep Learning
17.11 Radiomics and Dosiomics
17.12 Radiogenomics
17.13 Challenges Modelling Radiotherapy Response
17.14 Summary
17.15 Conclusions
References
18: Smart Adaptive Treatment Strategies
18.1 Introduction
18.2 Adaptive Treatment in Radiotherapy
18.3 What Knowledge Is Needed for ACT?
18.3.1 Clinical Data
18.3.2 Treatment Data
18.3.3 Imaging Data
18.3.4 Biological Data
18.4 How to Develop Outcome Models Using This Knowledge?
18.5 How to Optimize Adaptation?
18.5.1 Classical MDP/RL Learning
18.5.2 Deep MDP/RL Learning
18.6 ACT Example in Radiotherapy
18.7 Discussion and Recommendation
18.8 Conclusions
References
19: Artificial Intelligence in Clinical Trials
19.1 Introduction
19.1.1 Background on Clinical Trials in Oncology and Radiology
19.1.2 Clinical Trials as the Gold Standard for Clinical Practice
19.1.3 Why Do Clinical Trials Fail?
19.2 Types of Clinical Trial Design
19.2.1 Adaptive Clinical Trials
19.3 Artificial Intelligence and Clinical Trial Design
19.3.1 Need for Artificial Intelligence in Clinical Trial Design
19.3.2 The Multiple Roles of Artificial Intelligence (AI) in Clinical Trial Design
19.3.3 Challenges for Artificial Intelligence in Clinical Trial Design
19.3.4 Example Application of Artificial Intelligence in Trial Design (SMART)
19.4 Discussion and Recommendations
19.5 Conclusions
References
Index