FSE-Mamba: A novel Frequency-Spatial Entanglement Mamba model for retinal vessel segmentation

Pixel-wise segmentation of retinal vessels remains a substantial challenge in clinical and research settings. Recently, Mamba-based methods have gained significant attention due to their global receptive field and linear computational complexity. However, existing Mamba-based methods face two major limitations. Primarily, the causal constraints inherent in Mamba-based methods create directional biases that compromise their effectiveness in dense prediction tasks. Additionally, the intricate interwoven structure of retinal vessels and their high spatial similarity to surrounding tissues further exacerbate feature discrimination challenges. To address these limitations, we propose the Frequency-Spatial Entanglement Mamba (FSE-Mamba) network. Specifically, the Frequency-Spatial Coordinate Mamba (FSCM) addresses directional constraints in Mamba-based architectures through multi-scale axial attention while enhancing vascular discriminability via frequency-domain guided attention. The Multi-Scale Frequency Perception Module (MSFPM) lowers pixel similarity effects by capturing inter-frequency relationships, while the Dual-domain Selective Entanglement Attention (DSEA) module integrates features across different domains through entangled learning to enhance the model's comprehensive understanding and representation of dual-domain information. Extensive quantitative and qualitative experiments on four widely used retinal vessel segmentation datasets demonstrate that the proposed method exhibits significant competitiveness compared to existing methods. The code is made publicly available at https://github.com/Mrxutaosun/FSE-Mamba.