RetinexDA: Progressive Disentanglement Domain Adaptation for Unsupervised Cross-Modality Medical Image Segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Deep neural networks excel in medical image segmentation with similar data.
• Performance degrades significantly with differing imaging protocols, vendors, or physics.
• This limits the clinical deployment of AI models in diverse real-world scenarios.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation framework for medical image segmentation.
• To address performance degradation caused by domain shifts in clinical imaging data.
• To enhance the robustness and reliability of segmentation models across different environments.

Main Methods:

• Proposed RetinexDA framework decomposes medical images into domain-invariant structure and domain-specific appearance.
• Introduced Disentangled Knowledge Distillation (DKD) for semantic alignment between pixel and latent spaces.
• Developed a Bézier-curve domain bridging strategy for smooth transitions between domains.

Main Results:

• RetinexDA demonstrated superior performance compared to state-of-the-art unsupervised domain adaptation methods.
• The framework effectively preserves anatomical details while mitigating modality-dependent variations.
• Experiments on abdominal CT and cardiac MRI segmentation validated the approach's effectiveness.

Conclusions:

• RetinexDA offers a robust solution for unsupervised domain adaptation in medical image segmentation.
• The method shows significant potential for scalable and reliable deployment in clinical settings.
• Preserving structural information and aligning semantic features are key to cross-domain adaptation.