Adversarial Consistency for Single Domain Generalization in Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Organ segmentation is crucial for medical image analysis.
• Deep learning models often require retraining for new data domains.
• Domain Generalization (DG) aims to improve model adaptability.

Purpose of the Study:

• To develop a novel adversarial domain generalization method for organ segmentation.
• To enable generalization to unseen contrasts and scanner settings without retraining.
• To train models on a single domain by synthesizing new data distributions.

Main Methods:

• Proposed a single-domain adversarial domain generalization approach.
• Introduced an adversarial domain synthesizer (ADS) to create synthetic domains.
• Implemented a mutual information regularizer with patch-level contrastive learning for semantic consistency.

Main Results:

• The method demonstrated effective organ segmentation across unseen modalities, scanning protocols, and scanner sites.
• The adversarial domain synthesizer successfully generated plausible data distributions.
• The mutual information regularizer ensured semantic consistency in synthetic domains.

Conclusions:

• The proposed method offers a robust solution for organ segmentation generalization.
• Single-domain training with synthesized domains is a viable alternative to multi-domain training.
• This approach significantly reduces the need for extensive retraining of deep learning models in medical imaging.