EM-Mamba: An Edge-Mix Enhanced Long-Range Sequential Modeling Mamba for Kidney Segmentation in CT Scans

Manual delineation of the kidney region in CT scans is a laborious process. Although existing models have achieved remarkable success in the realm of automated segmentation, they still encounter challenges in effectively modeling both the global and local information. To address this, we introduce Edge-Mix enhanced Mamba (EM-Mamba) for kidney segmentation, which is designed to capture global and local information from multi-scales. EM-Mamba leverages SegMamba as its backbone, utilizing Mamba’s efficiency in extracting long-range dependencies. To further enhance the model’s integration of local and global information, we introduce the Edge-Mix module, which captures edge information from the shallow layers of the network during the encoding phase and merges it with multi-scale features during the decoding phase. Additionally, a Multi-Scale Convolution (MSC) module is designed to enhance the capability of extracting effective information from 3D medical images, achieving a fusion of rich local and global information across multiple scales. The innovation of the EM-Mamba model lies in its integration of detailed and global information, as well as its robust feature extraction ability in 3D medical image analysis, offering a new solution for precise kidney segmentation. The experimental results show that our framework achieving a Dice score of $\mathbf{9 1 . 9 5}$ % on the KiTS21 dataset and 86.35 % Dice score on our proprietary dataset called Kidney Parenchyma Segmentation (KPS).