AADG

AADG enhances domain generalization of convolutional neural networks for medical image segmentation by automatically learning and applying data augmentation policies that generate diverse augmented domains to reduce source–target domain gaps.

Key Features:

• Data Manipulation-Based Domain Generalization: Manipulates data augmentation policies within a defined search space to generate diverse novel domains and improve CNN robustness to domain shifts.
• Diverse Domain Generation: Uses a proxy task to maximize diversity among multiple augmented domains, quantified by the Sinkhorn distance in a unit-sphere embedding space.
• Adversarial Training and Deep Reinforcement Learning: Integrates adversarial training and deep reinforcement learning to efficiently search for and optimize augmentation policies.
• Model-Agnostic Policies: Produces augmentation policies that are empirically validated to transfer across different models without significant loss of performance.

Scientific Applications:

• Retinal Vessel Segmentation: Demonstrated generalization capabilities on four public fundus image datasets.
• Optic Disc and Cup (OD/OC) Segmentation: Validated across four datasets for OD/OC segmentation tasks.
• Retinal Lesion Segmentation: Assessed on three datasets for lesion segmentation performance.
• Retinal Vasculature Segmentation on OCTA: Evaluated on two Optical Coherence Tomography Angiography (OCTA) datasets to confirm cross-modality generalization.

Methodology:

Computational methods include automated search of data augmentation policies via deep reinforcement learning, adversarial training to challenge models, a proxy task that maximizes domain diversity measured by the Sinkhorn distance in a unit-sphere space, and application of the learned augmentations to CNN segmentation models.