Unified and Semantically Grounded Domain Adaptation for Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Unsupervised domain adaptation for medical image segmentation faces challenges due to distinct source-accessible and source-free settings.
• Existing methods often lack a structured approach to anatomical knowledge generalization across domains.
• This divergence necessitates a unified framework for robust and consistent adaptation.

Purpose of the Study:

• To develop a unified, semantically grounded framework for unsupervised domain adaptation in medical image segmentation.
• To enable adaptation in both source-accessible and source-free settings with a single model architecture.
• To provide an interpretable and anatomically informed solution for cross-domain medical image analysis.

Main Methods:

• Learned a domain-agnostic probabilistic manifold representing anatomical regularities.
• Interpreted image content as a canonical anatomy retrieved from the manifold plus a spatial transformation.
• Employed a disentangled, interpretable formulation for semantically meaningful predictions.

Main Results:

• Achieved state-of-the-art performance on cardiac and abdominal datasets in both source-accessible and source-free settings.
• Demonstrated source-free performance closely matching source-accessible results, indicating high consistency.
• Showcased strong interpretability through manifold traversal for smooth shape manipulation.

Conclusions:

• The proposed framework offers a principled foundation for unified, anatomically informed, and interpretable domain adaptation in medical imaging.
• The model architecture inherently supports adaptability without explicit cross-domain alignment strategies.
• This approach bridges the gap between different domain adaptation settings, enhancing consistency and performance.