Mamba-DACR: Direction-Aware Connectivity-Refinement Network for Tiny Vessel Segmentation

Vessel segmentation is critical in medical analysis, yet it remains challenging due to the presence of faint, thin, and discontinuous structures. Existing architectures, such as CNNs and transformers, often struggle to capture essential vascular properties, particularly local directionality and longitudinal continuity, resulting in fragmented segmentation for sub-pixel vessels. To address these limitations, we introduce the Direction-Aware Connectivity-Refinement Network (Mamba-DACR), which explicitly incorporates vascular priors into a Mamba-based framework. Mamba-DACR introduces a Direction-Aware Strategy that jointly models spatial and frequency domains. Specifically, Learnable Directional Convolutions (LD-convs) adaptively align with local vessel orientations in the spatial domain, while Adaptive Gabor Filters (A-Gabors) maximize the spectral response in the frequency domain. Building upon these direction-aware features, we design a Dual-Spectrum Connectivity Mamba (DSC-Mamba) module to capture long-range dependencies and refine structural continuity. By leveraging both directional priors and global context modeling, DSC-Mamba effectively restores broken vessel segments and improves topological consistency. Extensive experiments on four public datasets demonstrate that our Mamba-DACR achieves competitive performance.