WMFU-Net:Multi-scale Wavelet Mamba network and Adaptive fusion for retinal vessel segmentation

Retinal vessel segmentation plays a critical role as a non-invasive auxiliary technique in the diagnosis of various ophthalmic diseases. However, fundus images are often characterized by low contrast and noise interference, and the complex structure of retinal vessels further increases the difficulty of accurate segmentation. Existing methods often struggle to accurately extract fine vessels. To address these challenges, we propose a multi-scale segmentation network named WMFU-Net. Specifically, we redesign the encoder of the network. First, leveraging the advantages of frequency-domain feature representation and the efficiency of the Mamba architecture, we propose a novel block termed Multi-scale Wavelet Mamba (MSWM). This block further enhances feature representation by fusing wavelet features across different stages, enabling the model to effectively capture multi-scale information. To further integrate features from the frequency and spatial domains, we design an Adaptive Feature Fusion (AFF) block, which employs dynamic weighting and feature interaction mechanisms to fuse frequency-domain information with spatial local details, thereby improving the model's feature representation capability and segmentation performance. Extensive experiments were conducted on three publicly available datasets — DRIVE, STARE, and CHASE_DB1 — on which the proposed method achieved sensitivity (Sen) of 87.19%, 88.28%, and 87.06%, respectively. These results demonstrate the effectiveness of the proposed framework for retinal vessel segmentation.