RobustEMD: Domain robust matching for cross-domain few-shot medical image segmentation

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Few-shot medical image segmentation (FSMIS) models typically perform poorly when data comes from different domains (e.g., varying modalities, institutions, or equipment).
• Clinical applications require models that can generalize across these diverse medical imaging data domains.

Purpose of the Study:

• To introduce Cross-domain Few-shot Medical Image Segmentation (CD-FSMIS) and propose a novel RobustEMD matching mechanism.
• To enhance the cross-domain generalization capability of medical image segmentation models.

Main Methods:

• Developed a RobustEMD matching mechanism utilizing Earth Mover's Distance (EMD).
• Incorporated a channel-wise feature decomposition strategy to divide features into local nodes.
• Implemented a texture structure aware weights generation method using Sobel-based gradients to restrain domain-specific features.
• Employed a boundary-aware Hausdorff distance for transportation cost calculation.

Main Results:

• The proposed RobustEMD mechanism significantly improved performance in cross-modal, cross-sequence, and cross-institution segmentation scenarios.
• Ablation studies confirmed the contribution of each component of the RobustEMD mechanism to enhanced performance.
• The model demonstrated strong generalization capabilities in heterogeneous medical imaging environments.

Conclusions:

• The RobustEMD mechanism effectively addresses the challenge of domain shift in few-shot medical image segmentation.
• This approach offers a promising solution for real-world clinical applications requiring robust segmentation across diverse data sources.
• The method shows significant potential for advancing medical image analysis in heterogeneous environments.