Topology-Aware Retinal Vasculature Segmentation using Multi-Branch Dynamic Convolutions and Topological Loss

Retinal vessel segmentation remains challenging due to the complex topology and scale variations of vascular networks, particularly in preserving continuity of fine capillaries. This paper proposes a novel framework that jointly addresses these challenges through two key innovations: multi-branch dynamic snake convolutions (MB-DSConv) with deformable kernels that adaptively trace vascular structures, and a topological loss function enforcing structural continuity through connectivity constraints. The MB-DSConv architecture employs parallel pathways with thickness-aware feature fusion to capture multi-scale vascular patterns, while dynamically adjusting kernel shapes to vessel morphology. Complementing this, the topological loss maintains anatomical consistency by penalizing disconnected segments and misaligned endpoints. Comprehensive evaluations across retinal image datasets demonstrate superior performance compared to state-of-the-art methods, with enhanced preservation of vascular continuity and improved accuracy in thin vessel segmentation. Qualitative results show the framework's ability to maintain complex branching patterns and reduce fragmentation artifacts, particularly in low-contrast regions. This synergistic integration of adaptive feature learning and topological constraints provides an effective approach for medical image segmentation tasks requiring precise structural preservation.Clinical relevance- This work enables more accurate detection of microvascular abnormalities in early-stage diabetic retinopathy through improved continuity preservation of thin retinal vessels. The topology-aware segmentation results provide reliable quantitative measurements of vascular branching patterns and tortuosity, supporting objective assessment of disease progression in clinical practice.