Domain-Adaptive Retinal Vessel Segmentation for Unannotated Fundus Images

Abstract

Accurate retinal vessel segmentation supports quantitative vascular analysis in assessing ocular and systemic diseases. Yet, its clinical scalability is constrained by limited pixel-level annotations and domain shift across heterogeneous fundus datasets. Thus, this study proposes a domain-adaptive retinal vessel segmentation model (DA-VesselNet), a weakly supervised approach that transfers vessel-segmentation knowledge from annotated source datasets to the unannotated Retinal Fundus Multi-Disease Image Dataset (RFMiD). The model was trained on DRIVE, CHASE_DB1, and FIVES, and adapted to RFMiD. The ResNet50 encoder with an attention-gated U-Net decoder, confidence-aware pseudo-label supervision, and perturbation consistency regularisation were used for the training process. Results on the held-out CHASE_DB1 indicated that DA-VesselNet achieved a Dice score of 0.5778, an Intersection over Union (IoU) of 0.4082, and an Area Under the Curve (AUC) of 0.9518. On 200 held-out FIVES test images with different pathological features, it achieved a Dice score of 0.7447 and an AUC of 0.9733, outperforming the source-only baseline model. To assess adaptation independently of source-adjacent data, the model was further tested on STARE and HRF, two domains excluded entirely from source training. Adaptation improved Dice by 0.0469 on STARE and 0.0056 on HRF with AUC gains of 0.0140 and 0.0121, respectively. Ablation analysis identified source-anchored supervision as the dominant contributor to performance. The adapted model was subsequently applied to generate vessel pseudo-labels for the RFMiD target domain, providing a structural resource for future vessel-informed analysis. These findings demonstrate that DA-VesselNet offers a scalable solution for creating clinically relevant pseudo-labelled vessels in fundus imaging with limited annotations.