Enhancing Medical Image Segmentation with Mamba and UNet++

Medical image segmentation is predominantly achieved with U-Net architectures based on Convolutional Neural Networks (CNNs). However, U-Net has two primary limitations. First, CNNs are constrained in modeling long-range dependencies, a limitation that is partially addressed by transformers, which face challenges due to their quadratic computational complexity. Second, there is a semantic gap in U-Net between feature maps in the encoder and decoder, especially between shallow and deep layers. To address these issues, we propose Mamba-UNet++, which alleviates the limited receptive field using a Visual State Space Duality (VSSD) vision block based on the improved Mamba2 VSS block. To bridge the semantic gap, Mamba-UNet++ replaces U-Net's direct skip connections with UNet++ dense skip connections and incorporates deep supervision during training. Extensive experiments on three datasets across different modalities show that Mamba-UNet++ outperforms competing methods, as evidenced by metrics such as DSC and HD95.