High-Order Interaction and Low-Order Parallelization of Features Fusion with Novel Mamba-UNet Architecture for Medical Image Segmentation

Medical image segmentation is an essential method for computer-aided diagnosis. Although image segmentation models based on convolutional neural networks (CNNs) and vision transformers (ViTs) have achieved significant advancements, CNNs struggle to effectively capture long-range dependencies, while ViTs face limitations in local information extraction and are hindered by quadratic computational complexity. Recently, their inherent issues have been successfully addressed by the state-space models in Mamba and 2D-selective-scan in Vision Mamba. However, the presence of noise and excessive redundant information in medical images limits the practicality of these methods. To address these challenges, we propose a highly effective and accurate high-low-order feature fusion visual state space module, named HL-VSS. This module primarily consists of two core components: multi-scale spatial convolution and high-low-order feature fusion (HLFF). The former component preliminarily suppresses noise and captures multi-scale feature information from medical images, accurately extracting edge and detail features for the fusion component. The latter processes these features, further reducing redundant information through high-order interaction with 2D-selective-scan, and fuses the local features obtained by low-order parallel Mamba, ultimately extracting deeper medical image features. We incorporate HL-VSS into a U-shaped architecture, named high-low-order feature fusion visual Mamba UNet (V-UNet). Comparison experiments and ablation studies are conducted on four publicly available medical image datasets to validate the strong competitiveness of V-UNet in medical image segmentation tasks. The code is available at https://github.com/ai-dqh0106N-UNet_Code.