UN-SAM: Domain-Adaptive Self-Prompt Segmentation for Universal Nuclei Images

This repository is an official PyTorch implementation of the paper "UN-SAM: Domain-Adaptive Self-Prompt Segmentation for Universal Nuclei Images" [paper] accepted by Medical Image Analysis.

Dependencies

• Python 3.10
• PyTorch >= 1.10.0
• albumentations 1.5.2
• monai 1.3.0
• pytorch_lightning 1.1.0

Code

Clone this repository into any place you want.

git clone https://github.com/CUHK-AIM-Group/UNSAM.git
cd UNSAM
mkdir data; mkdir pretrain;

Quickstart

• Train the UN-SAM with the default settings:

python train.py --domain_num $NUMBER OF DOMAINS$ --size $B$ --sam_pretrain pretrain/$SAM CHECKPOINT$

Data Preparation

The structure is as follows.

UN-SAM
├── data
│   ├── DSB-2018
|     ├── data_split.json
│   ├── MoNuSeg
|     ├── data_split.json
│   ├── TNBC
|     ├── data_split.json
│   ├── image_1024
│     ├── DSB_0000000.png
│     ├── MoNuSeg_0000000.png
│     ├── TNBC_0000000.png
|     ├── ...
|   ├── mask_1024
│     ├── DSB_0000000.png
│     ├── MoNuSeg_0000000.png
│     ├── TNBC_0000000.png
|     ├── ...   

Pre-trained Model Zoo

We provide all pre-trained models here.

Size	Domains	Checkpoints
UN-SAM-B	DSB-2018 + MoNuSeg + TNBC	Google Drive
UN-SAM-L	DSB-2018 + MoNuSeg + TNBC	TBA
UN-SAM-H	DSB-2018 + MoNuSeg + TNBC	TBA

Cite

If you find our work useful in your research or publication, please cite our work:

@article{chen2025sam,
  title={UN-SAM: Domain-adaptive self-prompt segmentation for universal nuclei images},
  author={Chen, Zhen and Xu, Qing and Liu, Xinyu and Yuan, Yixuan},
  journal={Medical Image Analysis},
  pages={103607},
  year={2025},
  publisher={Elsevier}
}

Acknowledgements

• SAM