AV-MapNet: geodesic vessel tokenization and sequential refinement for retinal vessel segmentation and artery–vein classification

Accurate retinal vessel segmentation and artery–vein (AV) classification are essential for computer-aided diagnosis of cardiovascular and ophthalmic diseases, yet most existing methods treat these tasks separately and rely on heuristic post-processing, limiting robustness in complex vascular structures and pathological conditions. In this work, we propose AV-MapNet, a topology-aware and computationally efficient sequential deep learning framework that jointly performs vessel segmentation and AV classification in an end-to-end manner. The architecture incorporates lightweight mini encoder–decoder refinement modules as learned structural regularizers to enforce vascular continuity and topological consistency, while dilated convolutions, multi-scale feature extraction, and Gaussian-weighted refinement enhance sensitivity to thin vessels and preserve anatomical details. Extensive experiments on multiple public retinal datasets demonstrate that AV-MapNet achieves state-of-the-art performance, with segmentation accuracies of 98.3%, 98.6%, and 98.5% on RITE, DRIVE, and HRF datasets, respectively, and Dice scores of 87.01%, 88.20%, and 87.60%. For AV classification, the model achieves accuracies of 95.2% on LES-AV, 96.6% on AV-DRIVE, and 97.0% on RITE, with consistently high Dice scores and strong precision, sensitivity, and specificity. These results confirm the robustness of the proposed framework in handling pathological variability and complex vessel crossings, while maintaining low computational complexity, making AV-MapNet suitable for real-time clinical and teleophthalmology applications.