Contrastive Graph Modeling for Cross-Domain Few-Shot Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Cross-domain few-shot medical image segmentation (CD-FSMIS) is crucial for data-scarce medical applications.
• Existing methods often sacrifice source-domain accuracy for generalization by removing domain-specific information.

Purpose of the Study:

• To develop a novel framework, Contrastive Graph Modeling (C-Graph), that enhances CD-FSMIS performance.
• To address the limitations of existing methods by preserving domain-specific information and improving both cross-domain and source-domain accuracy.

Main Methods:

• Representing image features as graphs with pixels as nodes and semantic affinities as edges.
• Introducing a Structural Prior Graph (SPG) layer for capturing and transferring target-category node dependencies.
• Implementing a Subgraph Matching Decoding (SMD) mechanism for guided prediction using semantic relations.
• Designing a Confusion-minimizing Node Contrast (CNC) loss to reduce node ambiguity and enhance discriminability.

Main Results:

• C-Graph significantly outperforms existing CD-FSMIS approaches on multiple cross-domain benchmarks.
• The method achieves state-of-the-art performance in cross-domain segmentation.
• Strong segmentation accuracy is maintained on the source domain.

Conclusions:

• C-Graph offers a data-efficient and effective solution for CD-FSMIS.
• The framework successfully leverages structural consistency as a domain-transferable prior.
• C-Graph advances the field by improving generalization without compromising source-domain performance.