MSA-VMUNet: multi-scale augmented visual Mamba UNet for medical image segmentation

Abstract

Recent advances in Vision Mamba have shown strong capability in modeling long-range dependencies for image segmentation. However, the existing Mamba-based model, such as VMUNet, still faces limitations in multiscale feature extraction and training stability. To address these issues, we propose MSA-VMUNet, a multi-scale enhanced variant of VMUNet. The proposed model integrates a Multi-scale Feature Enhancement Block (MFEB) within each encoder and decoder stage, and employs Atrous Spatial Pyramid Pooling (ASPP) at the encoder output to better capture contextual information across varying object scales. Moreover, we introduce a learnable output scaling factor in the VSS block to improve optimization stability. Extensive experiments on the ISIC17 and ISIC18 datasets demonstrate that MSA-VMUNet consistently outperforms the original VMUNet and other state-of-the-art methods, particularly in segmenting small lesions and images with complex backgrounds.