RCDAN: A novel network for retinal vessel segmentation with rotational convolution and dynamic attention

Retinal vessel segmentation is critical for diagnosing ocular and systemic diseases, yet its complex morphology, characterized by varying thickness and curvilinear structures, presents significant difficulties. While deep learning has substantially advanced the field of image segmentation, the task remains difficult due to the intricate vessel geometry, the high computational cost of Vision Transformer (ViT) self-attention, and the limited multi-scale feature integration in existing methods. To address these challenges, we propose a Rotational Convolutional Dynamic Attention Network (RCDAN), which incorporates three novel modules, namely the Adaptive Rotated Inception Depthwise Convolution (ARIDC) module, the Dynamic Selective Single-Head Vision Transformer (DS-SHViT) module, and the Unified Attention (UA) module. Results from experiments on the DRIVE and CHASEDB1 datasets, along with cross-dataset evaluation on the unseen STARE dataset, demonstrate that RCDAN achieves a favorable balance between segmentation accuracy and computational efficiency. Notably, RCDAN operates at a real-time inference speed of 215 FPS with only 10.82M parameters, facilitating its integration into real-world and portable clinical workflows. Moreover, qualitative and topological analysis reveals improved delineation of fine and low-contrast vessels, underscoring the model's robustness and sensitivity to challenging anatomical structures.