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

Recent efforts have adapted cross-domain few-shot segmentation (CD-FSS) techniques to medical imaging, demonstrating initial success in bridging modality or institutional gaps. However, these approaches remain limited by two critical issues. First, they rely on unrefined prototype-based coarse predictions, which become highly unreliable under domain shift and propagate errors into matching stages. Second, these approaches suffer from unimodal reliance. Spatial-only methods are highly sensitive to domain-specific appearance variations, whereas frequency-only methods often compromise fine-grained geometric fidelity due to spectral abstraction. To overcome these issues, we propose the Frequency-Spatial Collaborative Matching Network (FSCMNet). Specifically, FSCMNet refines the coarse query mask via pixel-level supervised contrastive learning to improve feature discriminability. It further performs collaborative matching by jointly aligning support-query features in both the mid-frequency domain and the spatial domain. Finally, a bidirectional cross-attention fusion module enables mutual enhancement between the two streams. Experiments on three cross-domain benchmarks show FSCMNet achieves state-of-the-art performance, significantly outperforming existing methods. The source code of this work is available at https://github.com/Mentholatum/FSCM-Net.