Abstract
Existing few-shot medical image segmentation (FSMIS) models fail to address apractical issue in medical imaging: the domain shift caused by differentimaging techniques, which limits the applicability to current FSMIS tasks. Toovercome this limitation, we focus on the cross-domain few-shot medical imagesegmentation (CD-FSMIS) task, aiming to develop a generalized model capable ofadapting to a broader range of medical image segmentation scenarios withlimited labeled data from the novel target domain. Inspired by thecharacteristics of frequency domain similarity across different domains, wepropose a Frequency-aware Matching Network (FAMNet), which includes two keycomponents: a Frequency-aware Matching (FAM) module and a Multi-Spectral Fusion(MSF) module. The FAM module tackles two problems during the meta-learningphase: 1) intra-domain variance caused by the inherent support-query bias, dueto the different appearances of organs and lesions, and 2) inter-domainvariance caused by different medical imaging techniques. Additionally, wedesign an MSF module to integrate the different frequency features decoupled bythe FAM module, and further mitigate the impact of inter-domain variance on themodel's segmentation performance. Combining these two modules, our FAMNetsurpasses existing FSMIS models and Cross-domain Few-shot Semantic Segmentationmodels on three cross-domain datasets, achieving state-of-the-art performancein the CD-FSMIS task.