Aizimaiti Xiaokaiti

Introduction: Retinal vessel segmentation is difficult because fundus images contain branching structures, scale variation, uneven illumination, lesions, and weak vessel boundaries. Methods: We propose BGAFS-UNet, combining Boundary Perception Module (BPM), Lightweight Wide Gated Attention (LWGA) with Dynamic Channel Allocation (DCA) and Adaptive Branch Fusion (ABF), Cross-Stage Fusion (CSF), Contextual Mixing (ContMix), and Attention-Guided Dynamic Sampling (AGDS). We evaluate three prespecified image-disjoint runs on DRIVE, STARE, and CHASE and use image-level paired tests with Holm correction on DRIVE. Results: The full model obtains F1 scores of 80.61% on DRIVE, 76.15% on STARE, and 79.53% on CHASE. On DRIVE, corrected improvements are metric- and vessel-scale-specific: boundary F-score, clDice, and large-vessel Dice are significant, whereas Dice and small-vessel Dice are not. Discussion: BGAFS-UNet is parameter-efficient but not unconditionally more efficient, with higher FLOPs and latency than U-Net; the findings support structure-aware segmentation while motivating broader cross-device validation and optimization. The code and reproducibility materials remain available at https://gitee.com/Zhangguoce/bgafs-net.git , fixed at commit 477b217990f68d43d5c1c8f3a818d8cf2c8af384.