MamUNet: Mamba Enhances Vessel Segmentation in NIR-II Fluorescence Imaging

The Transformer architecture achieves excellent global context understanding of sequential data through the self-attention mechanism, but it faces computational challenges when modeling long-range dependencies. Mamba, based on the state space model (SSM), has emerged as a notable approach for effectively addressing these long-range dependencies, significantly improving computational and storage efficiency. Therefore, we propose MamUNet, a new method that combines the powerful modeling capabilities of the Mamba model with the strengths of the UNet architecture for complex vessel segmentation tasks in NIR-II fluorescence images. The encoder employs the classic network ResNet to learn local low-level features, followed by Mamba for global feature learning to capture more contextual information. We conducted experiments on a dataset of human brain NIR-II fluorescence vessel images and three publicly available retinal vessel datasets. The results demonstrate that our method achieves more accurate vessel segmentation with lower computational cost, outperforming other comparison methods.