Robust Few-Shot segmentation for multi-modal medical images using Fourier-Guided style fusion and prototype alignment

Document Type

Article

Publication Date

3-1-2026

Abstract

Few-shot semantic segmentation (FSS) in multi-modal medical imaging remains a significant challenge due to the scarcity of labeled samples and domain gaps induced by varying imaging devices, protocols, and anatomical structures. To tackle these issues, we propose a robust framework that leverages Fourier-guided style fusion and adaptive prototype alignment to improve cross-domain generalization. Specifically, we introduce a frequency-aware style transfer mechanism that decomposes source images into spectral components and infuses target-domain style in the frequency domain, effectively narrowing appearance discrepancies. To preserve semantic coherence and spatial structure, a CutMix-based augmentation is further applied. Additionally, a prototype alignment module, enhanced by cross-attentional feature interaction, facilitates task-specific knowledge transfer between support and query samples. The proposed model employs a self-supervised transformer backbone pre-trained for domain-invariant representation learning. Extensive experiments conducted on five publicly available datasets—ISIC2018, Chest-X, Kvasir, WBC, and SegPC-2021—demonstrate that our approach consistently surpasses existing FSS baselines under both 1-shot and 5-shot settings, with notable gains on complex and noisy modalities. These results highlight the potential of our framework for practical deployment in real-world biomedical segmentation tasks.

Keywords

Cross-domain learning, Few-shot semantic segmentation, Fourier-guided style fusion, Multi-modal medical imaging, Prototype alignment

Publication Title

Biomedical Signal Processing and Control

ISSN

1746-8094

DOI

10.1016/j.bspc.2025.108986

Funders

Project 333 of Jiangsu Province, Suzhou Laboratory of Data Innovation and Application, Jiangnan University, Program for Jiangsu Provincial Excellent Scientific and Technological Innovation Team, Changshu Key Laboratory of Medical ArtificialIntelligence ( CS202314, CYZ202301), Jiangsu Province Excellent Teacher Support Program under the “Qing Lan Project” in Higher Education Institutions (27), Jiangsu Province Vocational Education (31), Universiti Malaya (FP007-2023, MG015-2024), Natural Science Foundation of Shandong Province (ZR2024MF134), Engineering Technology Research and Development Center for Higher Vocational Colleges in Jiangsu Province (11)

Volume

113

Publisher

Elsevier Ltd

Share

COinS