VesselMamba: 3D vessel segmentation in CTA images using Mamba with enhanced Spatial-Channel Attention

3D vessel segmentation in Computed Tomography Angiography (CTA) is crucial yet challenging due to the complex, multi-scale, and elongated branching structure of human vasculature. Accurate modeling requires capturing both long-range dependencies and multi-scale information inherent in vascular networks. While deep neural networks like CNNs and Vision Transformers (ViTs) have demonstrated progress, they often face challenges balancing global receptive field capture with computational efficiency. To address these limitations, we propose VesselMamba, a novel 3D vessel segmentation framework based on Mamba, an approach for modeling long-range dependencies with linear complexity. VesselMamba integrates parallel Mamba blocks in the encoder to efficiently capture vascular continuity and long-range dependencies. Additionally, the encoder is enhanced with a Spatial-Channel Attention with Spatial Pyramid Pooling (SCASPP) module to effectively model multi-scale information and optimize the integration of global and local features, significantly improving segmentation precision. Furthermore, a composite loss function that combines the clDice loss with traditional cross-entropy and Dice losses is employed to improve the connectivity of segmented vessels. This reduces fragmentation and artifacts, leading to more reliable segmentations. Comprehensive ablation studies on private and public datasets demonstrate the complementary nature and effectiveness of the proposed modules. Experimental results show that VesselMamba achieves state-of-the-art performance in CTA vessel segmentation tasks, outperforming existing methods and providing a robust tool for clinical diagnosis and research.