RCDAN: A Novel Network for Retinal Vessel Segmentation with Rotational Convolution and Dynamic Attention

Retinal vessel segmentation (RVS) is a fundamental task in fundus image analysis and is critical for diagnosing various ocular and systemic diseases. Retinal vessels, as key biomarkers, exhibit complex morphologies such as varying thicknesses and curvilinear structures, making segmentation particularly challenging. While deep learning has advanced RVS, three major challenges remain: 1) convolutional neural networks (CNNs) have limited adaptability to the intricate shapes of vessels; 2) multi-head self-attention in Vision Transformers (ViTs) is computationally expensive and prone to redundancy, particularly on small datasets; 3) existing attention mechanisms lack the ability to effectively model and integrate multi-scale features.To address these challenges, we propose the Rotational Convolutional Dynamic Attention Network (RCDAN), incorporating: 1) Adaptive Rotated Inception Depthwise Convolution (ARIDC) for capturing multi-scale, multi-directional vessel features; 2) Dynamic Selective SHViT (DS-SHViT) for reducing redundancy and efficiently modeling global context; 3) Unified Attention Module (UAM) for enhancing multi-scale feature fusion. Experiments on the DRIVE and CHASEDB1 datasets show that RCDAN achieves superior performance compared to state-of-the-art methods.