Abstract
Retinal vessel segmentation is a critical step in computer-aided diagnosis of retinal images, providing rich information of vascular morphology. However, existing works primarily focus on achieving high overall segmentation accuracy, often overlooking the errors in the topological structure that lead to disconnected vessels. In order to address these issues, a topological coherence preserving retinal vessel segmentation method is proposed. We adopt a two-stage model by cascading a refinement step after the typical segmentation network to enhance retinal vessel connectivity. A topology-preserving loss is then introduced to guide the refinement network to pay more attention to key regions with topological differences, allowing it to learn the correct topology structure and maximize the preservation of retinal vessel tree connectivity. Experimental results on the DRIVE, STARE, and CHASE_DB1 datasets demonstrate the effectiveness of the proposed method.