SCM-UNet: Spatial-channel Mamba UNet for medical image segmentation

Abstract

Medical image segmentation plays a critical role in ensuring accurate diagnosis and treatment planning. Despite significant advances in segmentation models based on Convolutional Neural Networks (CNNs) and Transformers, challenges still exist in modeling long-range dependencies and managing computational complexity effectively. To address these challenges, we propose a novel architecture for medical image segmentation, called Spatial-Channel Mamba-UNet (SCM-UNet). This model incorporates the Structured Space Model (SSM) to capture remote dependencies with linear computational complexity, while also leveraging CNNs for local feature extraction. Additionally, we introduce the Spatial-Channel Attention Bridge (SCAB) module, which facilitates multi-scale feature fusion and enhances the model's expressiveness. Comprehensive experimental evaluations on five public benchmark datasets demonstrate that SCM-UNet achieves state-of-the-art (SOTA) performance. Specifically, for skin lesion segmentation, it obtains a mean Intersection over Union (mIoU) of 81.02% on the ISIC 2017 dataset and 81.88% on the ISIC 2018 dataset. To validate its generalizability, SCM-UNet was also evaluated on polyp (Kvasir-SEG, ColonDB) and breast ultrasound (BUSI) segmentation tasks, where it consistently outperformed existing methods, achieving top-ranking mIoU scores of 83.86%, 63.67%, and 69.03%, respectively. Overall, SCM-UNet effectively balances long-range dependency modeling with computational efficiency, offering a robust and versatile solution for various medical image segmentation tasks. This approach represents a promising direction for future research in improving both inference efficiency and accuracy.