DCSF-Net: Thin-Vessel-Aware Dynamic Cross-Scale Fusion for Retinal Vessel Segmentation

Abstract

Accurate retinal vessel segmentation is an important foundation for assisted screening and quantitative analysis of ophthalmic diseases and systemic diseases. However, the edge and high-frequency responses of thin, low-contrast vessels are easily attenuated during successive convolution operations and multilevel downsampling. In addition, conventional U-Net skip connections usually fuse encoder and decoder features directly, ignoring semantic differences and reliability across channels, which may introduce background textures and irrelevant responses into the decoder. To address these issues, this paper proposes a retinal vessel segmentation network named DCSF-Net. At the encoder side, a Shallow Edge-Guided Residual Enhancement module (SA-EGRE) is designed to selectively modulate local high-frequency residuals through edge confidence, thereby alleviating feature degradation at weak vessel boundaries. At the skip connections, a Dynamic Cross-Scale Channel Fusion (DCCF) module is introduced to implement dynamic cross-scale fusion by adaptively calibrating skip features according to the current decoder semantics and feature discrepancies. Experiments on the public DRIVE, STARE, and CHASE_DB1 datasets demonstrate favorable segmentation performance and improved vascular centerline consistency, suggesting that the proposed method helps to better preserve thin vessel branches.