SegUMamba: Integrating Mamba with U_net for Medical Image Segmentation

In medical image segmentation, capturing long-range dependencies is critical for precise segmentation. Convolutional Neural Networks(CNNs)excel at local features, while Transformers comprehend global context, but both struggle with long-range dependencies. Mamba addresses this with its selection mechanism and dynamic matrix. By integrating Mamba with U-Net, we developed the SegUMamba model, which effectively captures local lesion details as well as distant interactions between lesions and surrounding skin. The encoder-decoder architecture, built on VMamba and utilizing skip connections, merges shallow and deep-level information, enhancing segmentation accuracy. We designed a flexible integration mechanism that enables feature exchange between the encoder and decoder while restoring features within the decoder, achieving efficient and lossless multiplexing. Experiments on the ISIC2018 Task1 dataset show that, with the same system configuration and hyperparameter settings, our model significantly outperforms others in accuracy, completeness and efficiency, making it promising for complex medical image segmentation tasks.