<p>Accurate artery–vein (A/V) classification in retinal fundus images is a challenging task in automated vascular analysis. Fundus imaging provides a non-invasive view of the retina’s vascular network, where abnormalities can indicate sight-threatening diseases such as glaucoma, arteriosclerosis, and age-related macular degeneration. Reliable A/V classification supports clinicians in early diagnosis, treatment planning, and risk assessment, yet traditional image analysis struggles with small vessel detection, vessel crossings, and low-contrast regions. To address these challenges, this study proposes Y-Net, a novel deep learning architecture specifically designed for artery–vein classification. Unlike conventional U-Net variants, Y-Net introduces dual decoders optimized separately for arteries and veins, a parallel multi-scale input branch to enhance thin-vessel detection, and DropBlock regularization to improve robustness under limited data. Experimental evaluations on the DRIVE and CHASE_DB1 datasets demonstrate that Y-Net consistently performs better than the present state-of-the-art models. Specifically, on the DRIVE dataset, Y-Net achieves a sensitivity of 0.85, specificity of 0.97, accuracy of 0.97, and an AUC of 0.97, while on the CHASE_DB1 dataset, it attains a sensitivity of 0.88, specificity of 0.98, accuracy of 0.95, and an AUC of 0.94. These results confirm Y-Net’s superiority in handling complex structures, vessel crossings, and low-contrast capillaries, while maintaining high specificity and accuracy. However, this study is limited to two publicly available datasets, without large-scale external validation, and future work will extend Y-Net to larger, multi-centre datasets and apply cross-validation to further establish its clinical reliability.</p>