Contents
Preface
About the Authors
Acknowledgments
1. Introduction
1.1 Current Status of Fundus Lesions
1.2 The Progress of Artificial Intelligence in Intelligent Medical Diagnosis
1.3 DNN and the Diagnosis of Ocular Fundus Lesions
1.4 The Main Content of This Book
2. Automated Analysis for Retinopathy of Prematurity by Deep Neural Networks
2.1 Methods
2.1.1 Image labeling
2.1.2 Deep neural networks
2.1.3 Expertsβ comparisons
2.1.4 Testing of the DeepROP system under clinical settings
2.2 Results
2.2.1 Large-scale ROP datasets
2.2.2 Performance of the model
2.2.3 Comparison of the model with human experts
2.2.4 DeepROP website and clinical test
2.3 Discussions
3. DeepROP: An Automated ROP Screening System
3.1 Retinopathy of Prematurity
3.2 Related Works
3.2.1 Traditional methods for diagnosis of ROP
3.2.2 Deep neural networks for diagnosis of ROP
3.3 Data and Methodology
3.3.1 Data
3.3.2 Methodology
3.4 Experimental Setup and Results
3.4.1 Experimental setup
3.4.2 Results
3.5 Clinical Application of DeepROP
3.5.1 System framework
3.5.2 System implementation
3.5.3 Clinical application
3.6 Conclusion
4. Diagnosis of Diabetic Retinopathy Using Deep Neural Networks
4.1 Introduction
4.2 Related Works
4.2.1 Related dataset
4.2.2 Traditional practice
4.2.3 Deep convolutional neural network approaches
4.3 Overview
4.4 Dataset Construction
4.5 Data Preprocessing and Augmentation
4.5.1 Preprocessing
4.5.1.1 Size normalization
4.5.1.2 Shape normalization
4.5.1.3 Color normalization
4.5.2 Augmentation
4.6 Model
4.7 Experiments
4.7.1 Experimental setup
4.7.2 Details of learning
4.7.3 Results
4.7.4 Visualization
4.8 Model Deployment and Clinical Evaluation
4.9 Conclusion
5. Automated Identification and Grading System of Diabetic Retinopathy Using Deep Neural Networks
5.1 Related Works
5.1.1 Traditional methods for DR diagnosis
5.1.2 Deep learning for DR diagnosis
5.2 Dataset
5.2.1 Materials
5.2.2 Grading standard
5.2.3 Manual grading
5.2.4 Preprocessing of retinal images
5.2.5 Performance comparison
5.2.6 Data augmentation
5.3 Model and Methodology
5.3.1 Aim and objective
5.3.2 Architecture and strategy of ensemble model
5.3.3 Ensemble strategy
5.3.4 Transfer learning at the first part
5.3.5 Design of customized Standard Deep Neural Network (SDNN) at the second part
5.4 Experiments
5.4.1 Configuration
5.4.2 Strategy
5.4.3 Metrics
5.4.4 Identification system
5.4.5 Grading system
5.4.5.1 Two alternative strategies
5.4.5.2 Four-class classification
5.4.6 Analysis of experiments
5.5 Discussion
5.6 DeepDR: An AI System for DR Diagnosis
5.7 Conclusion
6. Automated Segmentation of Macular Edema in OCT Using Deep Neural Networks
6.1 Introduction
6.2 Related Works
6.2.1 Traditional methods for segmenting macular edema
6.2.2 Deep neural networks for segmenting macular edema
6.3 Methodology
6.3.1 Atrous spatial pyramid pooling
6.3.1.1 Atrous convolution
6.3.1.2 Multiple scale features
6.3.2 Stochastic atrous spatial pyramid pooling
6.3.2.1 Multiple scale features with randomness
6.3.2.2 Model ensemble
6.4 Experimental Setup and Results
6.4.1 Experimental setup
6.4.1.1 Dataset
6.4.1.2 Configurations
6.4.1.3 Implementation
6.4.2 Results
6.4.2.1 Comparison with the state of the art
6.4.2.2 Stochastic atrous spatial pyramid pooling
6.4.2.3 Analytical experiment
6.4.2.4 Comparison on 3DIRCADb dataset
6.5 DeepOCT: An AI System for Macular Edema Lesion Segmentation
6.6 Conclusion
7. DeepUWF: An Automated Ultrawide-field Fundus Screening System via Deep Learning
7.1 Introduction
7.2 Related Works
7.2.1 Deep learning with traditional imaging technology
7.2.2 Deep learning with emerging UWF imaging technology
7.3 Dataset
7.3.1 Materials
7.3.2 Annotation standard
7.3.3 Annotation result
7.3.4 Preprocessing of UWF images
7.3.5 Data augmentation
7.4 Model and Methodology
7.4.1 Aim and objective
7.4.2 Strategy and framework
7.4.3 Selection of feature extractors
7.4.4 Design of customized classifiers
7.4.5 Performance metrics
7.5 Experiments
7.5.1 Experimental setup
7.5.2 Screening system
7.5.2.1 System design
7.5.2.2 Analysis of empirical
7.5.2.3 Generalization performance
7.5.3 Diagnostic system
7.5.3.1 Data partition
7.5.3.2 System design
7.5.3.3 Analysis of empirical
7.6 Discussion
7.6.1 Image optimization
7.6.2 Medical implication
7.6.3 Limitations
7.6.4 Future work
7.7 Conclusion
8. DeepUWF-Plus: Automatic Fundus Identification and Diagnosis System Based on Ultrawide-field Fundus Imaging
8.1 Related Works
8.1.1 Deep learning with traditional imaging technology
8.1.2 Deep learning with emerging UWF imaging technology
8.2 Dataset
8.2.1 Materials
8.2.2 Annotation standard
8.2.3 Annotation details
8.2.4 Data preprocessing and augmentation
8.3 Model and Methodology
8.3.1 Aim
8.3.2 Feature extractors
8.3.3 Custom classifications
8.3.4 Class imbalance
8.3.5 Evaluation metrics
8.4 Experiments
8.4.1 Experimental setup
8.4.2 Experimental design
8.4.3 Screening system
8.4.4 Sign identification system
8.4.4.1 Two-step classification strategy
8.4.4.2 One-step classification strategy
8.4.5 Disease diagnosis system
8.4.5.1 Two-step classification strategy
8.4.5.2 One-step classification strategy
8.4.6 Result
8.4.6.1 Screening system
8.4.6.2 Sign identification system
8.4.6.3 Disease diagnosis system
8.4.7 Comparative experiments
8.4.8 Generalization performance
8.5 Discussion
8.5.1 Comparison of the two strategies
8.5.2 Model performance
8.5.3 Medical implication
8.5.4 Research limitations
8.5.5 Future work
8.6 DeepUWF: An AI System for Multiple Fundus Diseases Diagnosis
8.7 Conclusion
Bibliography
Index