MGVSS-UNet: A Novel U-Net Architecture Integrating Multi-Scale Global Visual State Space for Medical Image Segmentation

Medical image segmentation plays an essential role in computer-aided diagnosis, supports physicians with fast and accurate information to make timely treatment decisions. However, the existing Mamba-based state space models (SSMs) lack a multi-scale feature learning mechanism in the encoder for medical image segmentation, and the reliance on specific selection mechanisms for modeling in the decoding stage makes it challenging to effectively integrate complex cross-layer semantic information. Therefore, we propose a Mamba UNet model, named MGVSS-UNet. This approach introduces the Multi-Scale Global Visual State Space Block (MSGVSS Block) in the encoder, which combines multi-scale convolution with Visual State Space (VSS) modeling to capture dynamic feature changes, while incorporating attention mechanisms to focus on key regions, thereby enhancing the model’s multi-scale and global contextual awareness. The decoder employs the Random Global-Local Context Aggregation (RGLCA) block, which effectively balances local details and global semantics, thereby enabling high-resolution semantic segmentation. Embedding a Depth-wise Separable Cross Fusion (DSCF Block) at the skip connection to assist the decoder in fully integrating the original and reconstructed features. Experiments on the benchmark datasets show that MGVSS-UNet outperforms existing methods significantly in terms of generalization ability and image segmentation performance, and suggest that such applications will become more widely used in the medical field in the future.