Generalized Retinal Artery/Vein Segmentation via Multi-Dataset Fine-Tuning and Pathology Subgroup Analysis

Abstract

Background: Automatic artery/vein (A/V) segmentation in color fundus photography underpins retinal biomarkers such as the arteriolar-to-venular ratio (AVR), yet single-dataset models generalize poorly across institutions and pathologies. Methods: Using pre-trained Recursive Refinement W-Net (RRWNet) weights as initialization, eight heterogeneous datasets were jointly fine-tuned under a single set of weights for 50 epochs: five A/V-labeled (Fundus-AVSeg, LES-AV, RITE, FIREFLY-Gen, PSEUDO_AV) and three vessel-only (DRIVE, CHASE_DB1, HRF). A vessel-only loss separation strategy applied only the vessel-channel loss to vessel-only datasets, preventing A/V metric dilution; post-processing combined field-of-view masking, connected-component denoising, and morphological refinement, outputting artery, vein, and vessel masks with crossing and junction maps. Results: Loss separation recovered validation DSC from 0.518 to 0.570 (AUC up to 0.959). On re-inference across four datasets (332 images), the single-weight model attained mean-AV DSC 0.657 ± 0.085 and AUC 0.976 ± 0.016; across normal, AMD, glaucoma, and DR subgroups, performance was statistically indistinguishable (Kruskal–Wallis, all p > 0.36), indicating consistent behavior across pathologies. Conclusions: Within a leakage-aware internal evaluation, the single-weight model produced stable A/V segmentation across the four datasets and across pathology subgroups, indicating preliminary internal consistency rather than proven cross-domain generalization. Because the evaluation reuses data seen during training and no external or patient-level held-out set was used, robustness claims are deferred; leakage-free external validation and patient-level re-evaluation are identified as the essential next steps.