MVM-UNet: multi branch convolutional vision Mamba UNet for medical image segmentation

State Space Models (SSMs), particularly the Mamba architecture, demonstrate significant potential in medical image segmentation due to the capabilities in modeling long-range dependencies. In the realm of medical image analysis, Vision Mamba UNet (VM-UNet) employs an asymmetric encoder-decoder architecture featuring Visual State Space blocks to capture extended spatial contextual information. However, VM-UNet encounters critical limitations: 1) Its restricted multi-scale modeling capacity during feature extraction impedes the effective fusion of local and global features, ultimately diminishing segmentation accuracy and robustness; 2) When processing complex lesion morphologies, certain boundary information loss occurs. This affects the method's applicability in precision-sensitive medical imaging scenarios. To address these limitations, we propose MVM-UNet, a multi-branch convolutional VM-UNet architecture. By extending VM-UNet, our model integrates a novel convolutional state-space module comprising parallel branches dedicated to capturing local and global features across varying receptive fields. The subsequent feature fusion operations enhance both representational capacity and generalization potential. Specifically, MVM-UNet achieves computational efficiency through depth-wise separable convolutional branches while utilizing heterogeneous depth-wise convolutions with multi-scale kernels for optimal texture characterization in complex lesion regions. Furthermore, the architecture incorporates a hybrid normalization scheme combining BatchNorm and LayerNorm with parametric ReLU activation, significantly improving training stability and feature discriminability. Comprehensive evaluations on the ISIC17 and ISIC18 skin lesion benchmarks, supplemented by a proprietary kidney tumor dataset, demonstrate the superiority of MVM-UNet over existing approaches. Quantitative analyses reveal significant improvements in DSC (increased by 0.37%) and mIoU (increased by 0.71%), statistically validating the framework's effectiveness in medical image segmentation tasks.