Retinal blood vessel morphology is crucial for diagnosing diseases like diabetic retinopathy, hypertension, and arteriosclerosis. However, manual annotation is labor-intensive, while automatic segmentation faces challenges due to complex backgrounds and fine vessel structures. To address this, we propose ME-Net, a U-Net-based model with two key improvements: (1) a Multi-scale Visual Attention Block (MSVA Block) that adaptively selects optimal feature scales to enhance global and local feature extraction, and (2) an Edge Branch (EB) with Efficient Gated Attention (EGA) to improve boundary segmentation. Trained using a dual-task learning framework with multiple loss functions, ME-Net achieves superior performance on DRIVE, STARE, and CHASE_DB1 datasets. Experimental results show consistent improvements in F1-score and Sensitivity over existing methods, demonstrating ME-Net’s accuracy and robustness in retinal vessel segmentation.