Long-range Sequential Modeling Mamba For 3D Abdominal Multi-Organ Segmentation

The automatic segmentation of abdominal organs in CT images is a critical task in medical image processing, aiding in improved diagnosis, treatment planning, and disease monitoring. However, due to blurred organ boundaries, varying sizes and shapes, and complex anatomical structures in abdominal CT images, the automatic segmentation of multiple organs remains challenging. To address these issues, this paper proposes a novel U-shaped model called MUX-Net, which combines state-space models with ConvNet architecture. We introduce the Mixed Vision Strategy Module (MambaUX) in the encoder stage, serving as the backbone of the encoder. By integrating the state-space model Mamba and large-kernel volumetric depthwise separable convolutions, we leverage the complementary advantages of different visual encoder strategies. This enhances the segmentation model’s ability to extract both local and global features in 3D images. The effectiveness of the hybrid approach largely owes to Mamba’s ability to capture long-term dependencies in sequential data and the large receptive field provided by the large convolutions. We evaluated the proposed method on a challenging public dataset for abdominal multi-organ segmentation: MICCAI Challenge 2022 AMOS. Experimental results show that MUX-Net achieved a Dice and NSD score of 0.8816 and 0.7699 on the AMOS2022 dataset, outperforming other mainstream segmentation methods.