Exploiting Cross-Scale Attention Transformer and Progressive Edge Refinement for Retinal Vessel Segmentation

Accurate retinal vessel segmentation is a crucial step in the clinical diagnosis and treatment of fundus diseases. Although many efforts have been presented to address the task, the segmentation performance in challenging regions (e.g., collateral vessels) is still not satisfactory, due to their thin morphology or the low contrast between foreground and background. In this work, we observe that an intrinsic appearance exists in the retinal image: among the dendritic vessels there are generous similar structures, e.g., the main and collateral vessels are all curvilinear, but they have noticeable scale differences. Based on this observation, we propose a novel cross-scale attention transformer (CAT) to encourage the segmentation effects in challenging regions. Specifically, CAT consumes features with different scales to produce their shared attention matrix, and then fully integrates the beneficial information between them. Such new attention architecture could explore the multi-scale idea more efficiently, thus realizing mutual learning of similar structures. In addition, a progressive edge refinement module (ERM) is designed to refine the edges of foreground and background in the segmentation results. Through the idea of edge decoupling, ERM could suppress the background feature near the blood vessels while enhancing the foreground feature, so as to segment vessels accurately. We conduct extensive experiments and discussions on DRIVE and CHASE_DB1 datasets to verify the proposed framework. Experimental results show that our method has great advantages in the Se metric, which are 0.88-7.26% and 0.81-7.11% higher than the state-of-the-art methods on DRIVE and CHASE_DB1, respectively. In addition, the proposed method also outperforms other methods with 0.17-2.06% in terms of the Dice metric on DRIVE.