AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Convolutional neural networks (CNNs) excel at medical image segmentation but suffer from domain gaps.
• Performance degradation occurs when models trained on one dataset are applied to another with different characteristics.

Purpose of the Study:

• To propose a domain generalization method, Automated Augmentation for Domain Generalization (AADG), to address the domain gap in medical image segmentation.
• To develop a framework that samples data augmentation policies to generate novel domains and diversify training sets.

Main Methods:

• AADG employs a novel proxy task to maximize diversity among augmented domains using Sinkhorn distance.
• Adversarial training and deep reinforcement learning are utilized for efficient policy search.
• The method was evaluated on 11 fundus image datasets and 2 OCTA datasets.

Main Results:

• AADG achieved state-of-the-art generalization performance in retinal vessel, optic disc/cup, and lesion segmentation.
• The method significantly outperformed existing approaches on multiple segmentation tasks.
• Learned augmentation policies demonstrated model-agnostic properties and good transferability to other models.

Conclusions:

• AADG effectively mitigates the domain gap issue in medical image segmentation.
• The proposed automated augmentation strategy enhances model robustness and generalization capabilities.
• AADG offers a promising solution for improving the reliability of deep learning models in diverse clinical settings.