Structure-aware learnable multi-scale attention for retinal vessel analysis

Accurate retinal vessel segmentation is critical for automated screening and monitoring of ocular and systemic diseases; however, existing deep segmentation models often struggle to preserve thin, low-contrast vessels. We propose a Learnable Multi-Scale Vesselness Attention (LMSA) module, a lightweight and architecture-agnostic component that introduces explicit vessel-aware priors into deep feature learning by adaptively selecting scale-space vesselness responses under a second-order structural prior and refining them through directional coherence modulation. LMSA aggregates multi-scale vesselness responses into a structurally aware representation and transforms them into a spatial attention mask that selectively enhances curvilinear vascular structures while suppressing background noise. Using U-Net as a reference backbone, an LMSA-enhanced network is constructed with an integrated multi-task learning (MTL) design that jointly supervises vessel segmentation, boundary delineation, and centerline extraction, thereby enforcing structural consistency. The generality of the framework is further demonstrated by integrating LMSA and MTL into multiple CNN segmentation architectures. Extensive experiments on three public benchmarks indicate that the proposed U-Net+LMSA+MTL model outperforms state-of-the-art baselines, achieving Dice scores of 85.39%, 85.00%, and 85.47%, with corresponding accuracies of 97.37%, 97.87%, and 97.90% on DRIVE, STARE, and CHASE_DB1, respectively. On a clinically acquired in-house dataset, the model maintains strong performance with a Dice score of 82.77% and an accuracy of 95.88%, supporting its potential for real-world clinical deployment. Collectively, these results demonstrate that the proposed framework effectively embeds vessel prominence and structural priors into deep networks, enabling more reliable, structurally consistent, and clinically meaningful retinal vessel analysis.