MUP-SAM: Multi-scale vision mamba UNet prompt generation for SAM in multi-organ medical image segmentation

Multi-organ medical segmentation is an important task in medical image analysis, which supports diagnosis, treatment and monitoring of diseases by accurately outlining the contours of tumors and organs. In recent years, the Segment Anything Model (SAM) has brought breakthroughs in medical segmentation tasks. However, it is inefficient for most medical SAM methods to perform segmentation relying on external prompts. To address the above issue, we propose a novel framework that leverages Multi-Scale Vision Mamba UNet as an auxiliary segmentation network to generate box prompts for SAM (MUP-SAM) in multi-organ medical image segmentation. The framework contains two key modules to further achieve the high accuracy: (1) the post-processing module is proposed to enhance the generated prompt of the Multi-Scale Vision Mamba UNet, which can provide the refined box prompts for SAM, and (2) the prediction fusion module is proposed to integrate the SAM segmentation result with the auxiliary segmentation network result effectively. Our proposed MUP-SAM method achieved 86.43% and 93.17% Dice score on the Synapse and ACDC datasets, respectively. The experimental results show that our method achieves automatic SAM segmentation and consistently outperforms existing methods, including other prompt-free SAM methods.