Tubular-aware mamba for accurate retinal vessel segmentation: preserving fine details and topological connectivity

Dangguo Shao, Rui Xu, Lei Ma, Sanli Yi

Abstract: Accurate retinal vessel segmentation is crucial for diagnosing ocular and systemic diseases. Current methods struggle to balance noise discrimination with detail preservation, especially for capillary detection. This paper introduces TA-Mamba, a novel framework combining the State Space Model (SSM) Mamba with tubular structure perception modules. TA-Mamba utilizes a high and low frequency attention Mamba block, tubular-aware gated convolution, serpentine spatial convolution, directional feature fusion, and a convolution-up block. Experimental results on DRIVE, CHASE_DB1, and STARE datasets demonstrate TA-Mamba’s superior performance, outperforming state-of-the-art methods in accuracy, continuity, and topological connectivity. This work represents a significant advancement in retinal vessel segmentation, effectively addressing challenges posed by low contrast and high noise in retinal images.