# Customized Segment Anything Model for Medical Image Segmentation

**Authors:** Kaidong Zhang, Dong Liu  
**arXiv:** [2304.13785](https://arxiv.org/abs/2304.13785)  
**Date:** 2023-04-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.13785v2/

## Key points

- SAMed is a novel solution for medical image segmentation based on the Segment Anything Model (SAM)
- SAMed utilizes a low-rank-based fine-tuning strategy and the AdamW optimizer for convergence and lower loss
- SAMed achieves impressive results with an 81.88 Dice Similarity Coefficient and 20.64 Hausdorff Distance on the Synapse multi-organ segmentation dataset
- SAMed's ability to perform semantic segmentation on medical images sets it apart from previous methods
- The deployment cost and storage requirements of SAMed are minimal due to updating only a small fraction of parameters during training

## Summary

In their technical report titled "Customized Segment Anything Model for Medical Image Segmentation," authors Kaidong Zhang and Dong Liu introduce SAMed, a novel solution for medical image segmentation. Unlike previous methods, SAMed is based on the Segment Anything Model (SAM), a large-scale image segmentation model, aiming to explore the customization of such models for medical image segmentation. The approach involves applying a low-rank-based (LoRA) fine-tuning strategy to the SAM image encoder, along with the prompt encoder and mask decoder, on labeled medical image segmentation datasets. By incorporating warmup fine-tuning and utilizing the AdamW optimizer, SAMed demonstrates successful convergence and lower loss compared to its predecessors. One key distinction of SAMed is its ability to perform semantic segmentation on medical images. The authors report that their trained SAMed model achieves impressive results with an 81.88 Dice Similarity Coefficient (DSC) and 20.64 Hausdorff Distance (HD) on the Synapse multi-organ segmentation dataset, putting it on par with state-of-the-art methods in the field. Extensive experiments were conducted to validate the effectiveness of this design. Furthermore, SAMed's deployment cost and storage requirements are minimal in practical usage due to only updating a small fraction of the SAM parameters during training. The code for implementing SAMed is openly available at https://github.com/hitachinsk/SAMed, enabling further exploration and application of this innovative approach in medical image analysis.