# Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation

**Authors:** Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, Yueming Jin  
**arXiv:** [2304.12620](https://arxiv.org/abs/2304.12620)  
**Date:** 2023-04-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.12620v7/

## Key points

- The Segment Anything Model (SAM) has shown impressive performance in image segmentation tasks and features a prompt-based interface.
- SAM falls short in medical image segmentation due to a lack of domain-specific medical knowledge, prompting the development of the Medical SAM Adapter (Med-SA).
- Med-SA integrates domain-specific medical knowledge into the segmentation model using Space-Depth Transpose (SD-Trans) for adapting 2D SAM to 3D medical images and Hyper-Prompting Adapter (HyP-Adpt) for prompt-conditioned adaptation.
- Evaluation experiments on 17 different medical image segmentation tasks showed that Med-SA outperformed several state-of-the-art methods while updating only 2% of the parameters.
- The research findings have been published under the title "Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation" and are available on GitHub at https://github.com/KidsWithTokens/Medical-SAM-Adapter.

## Summary

, , , , 

In the field of image segmentation, the Segment Anything Model (SAM) has garnered attention for its impressive performance in various segmentation tasks and its prompt-based interface. However, recent studies have revealed that SAM falls short in medical image segmentation due to a lack of domain-specific medical knowledge. This limitation prompted researchers to explore ways to enhance SAM's segmentation capabilities for medical images. To address this issue, a team of researchers led by Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin proposed the Medical SAM Adapter (Med-SA). Unlike traditional approaches that involve fine-tuning the SAM model, Med-SA integrates domain-specific medical knowledge into the segmentation model using a lightweight yet effective adaptation technique. The key innovations introduced in Med-SA include Space-Depth Transpose (SD-Trans) for adapting 2D SAM to 3D medical images and Hyper-Prompting Adapter (HyP-Adpt) for prompt-conditioned adaptation. The researchers conducted comprehensive evaluation experiments on 17 different medical image segmentation tasks spanning various image modalities. The results demonstrated that Med-SA outperformed several state-of-the-art medical image segmentation methods while updating only 2% of the parameters. This significant improvement showcases the effectiveness of incorporating domain-specific knowledge into the segmentation model. Overall, the Medical SAM Adapter presents a promising solution to enhance SAM's segmentation capabilities for medical images. The research findings have been published under the title "Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation" and are available for further exploration through their released code on GitHub at https://github.com/KidsWithTokens/Medical-SAM-Adapter.