Frequency-Spatial Collaborative Matching for Cross-Domain One-Shot Medical Image Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Cross-domain few-shot segmentation (CD-FSS) shows promise in medical imaging for bridging data gaps.
• Existing CD-FSS methods struggle with unreliable coarse predictions and unimodal reliance (spatial-only or frequency-only).
• Spatial methods are sensitive to appearance variations, while frequency methods may lose geometric details.

Purpose of the Study:

• To address the limitations of current CD-FSS techniques in medical imaging.
• To propose a novel network, FSCMNet, for robust and accurate cross-domain segmentation.
• To enhance feature discriminability and geometric fidelity in medical image segmentation.

Main Methods:

• Developed the Frequency-Spatial Collaborative Matching Network (FSCMNet).
• Employed pixel-level supervised contrastive learning to refine query masks and improve feature discriminability.
• Implemented collaborative matching by aligning features in both mid-frequency and spatial domains.
• Utilized a bidirectional cross-attention fusion module for stream enhancement.

Main Results:

• FSCMNet achieved state-of-the-art performance on three cross-domain benchmarks.
• The proposed method significantly outperformed existing CD-FSS approaches.
• The collaborative matching strategy effectively addressed domain shift challenges.

Conclusions:

• FSCMNet offers a significant advancement in cross-domain few-shot medical image segmentation.
• The integration of frequency and spatial information leads to more robust and accurate segmentation.
• The developed network overcomes critical limitations of prior unimodal and coarse prediction methods.