📌 This is an official PyTorch implementation of De-LightSAM: Modality-Decoupled Lightweight SAM for Generalizable Medical Segmentation
[2024.08.08] The pre-print paper has been uploaded!
[2024.08.07] Paper will be updated soon!
[2024.08.07] Code and model checkpoints are released!
git clone https://github.com/xq141839/De-LightSAM.git
cd De-LightSAM
conda create -n ESP python=3.10
conda activate ESP
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia
pip install albumentations==0.5.2
pip install pytorch_lightning==1.1.1
pip install monaiNote: Please refer to requirements.txt
The structure is as follows.
De-LightSAM
├── datasets
│ ├── image_1024
│ ├── ISIC_0000000.png
| ├── ...
| ├── mask_1024
│ ├── ISIC_0000000.png
| ├── ...
We provide all pre-trained models here.
| MA-Backbone | MC | Checkpoints |
|---|---|---|
| TinyViT | Dermoscopy | Link |
| TinyViT | X-ray | Link |
| TinyViT | Fundus | Link |
| TinyViT | Colonoscopy | Link |
| TinyViT | Ultrasound | Link |
| TinyViT | Microscopy | Link |
If you find this work helpful for your project, please consider citing the following paper:
@article{xu2024esp,
title={ESP-MedSAM: Efficient Self-Prompting SAM for Universal Domain-Generalized Medical Image Segmentation},
author={Xu, Qing and Li, Jiaxuan and He, Xiangjian and Liu, Ziyu and Chen, Zhen and Duan, Wenting and Li, Chenxin and He, Maggie M and Tesema, Fiseha B and Cheah, Wooi P and others},
journal={arXiv preprint arXiv:2407.14153},
year={2024}
}