HDA-SAMUNet: A Medical Image Segmentation Method Based on Vision Mamba UNet and MedSAM

Deep learning methods have been extensively explored for medical image segmentation. Although recent CNN--based and Transformer-based frameworks have advanced medical-image segmentation, they remain computationally burdensome and yield suboptimal boundary delineation, hindering their adoption in time-critical clinical workflows. We propose a novel medical image segmentation method to address these issues by integrating the Visual State Space model Vision Mamba UNet(VM-UNet) with the medical image foundation model MedSAM. First, we introduce a LD Block at the input stage and a LCC Block within the existing VM-UNet framework, forming a new sub-network named HDA-UNet (a Hybrid model combining Dense connectivity and Attention in VM-UNet). Then, we incorporate high-quality features extracted from MedSAM into the deep layers of the U-shaped network through dual parallel encoders, enhancing the representational capacity of the encoded features. Finally, the fused features are fed into a multi-scale decoder to generate accurate segmentation outputs, resulting in the complete architecture HDA-SAMUNet. We evaluate our method on several public datasets, including ISIC-2017 and ISIC2018. Empirical results show that HDA-SAMUNet surpasses VM-UNet by $0.89 \% \mathrm{mIoU}$, 0.55% DSC, and 0.26% Acc on ISIC2017, with similar gains on ISIC-2018, underscoring its potential to deliver more accurate and dependable medical-image segmentations in clinical practice.