FS_MedSAM2: Exploring the Potential of SAM2 for Few-Shot Medical Image Segmentation without Fine-tuning

arxiv

News

We provide example_data folder, which contains some example data for anyone who want to try this code.

Getting Strated

Data preparation

You can obtain the data by following steps, or directly obtain the data by Baidu Cloud (password: t9vu).

1. Follow SSL_ALPNet to create SSL_ALPNet project at /path/to/SSL_ALPNet;
2. Process Synapse-CT and CHAOS-MRI datasets follow SSL_ALPNet;
3. Place validation_wopred.py in the /path/to/SSL_ALPNet directory, update the saved_npz_path = '/path/to/saved_npz', and run it using either /path/to/SSL_ALPNet/test_ssl_abdominal_ct.sh or /path/to/SSL_ALPNet/test_ssl_abdominal_mri.sh.

Experiments Reproduction

1. Based on the official repository of SAM2, deploy SAM2 locally at /path/to/SAM2;
2. Move the files and folders from this repository, /sam2 and those under /notebooks, to the corresponding folder at /path/to/SAM2/sam2 and /path/to/SAM2/notebooks. e.g, move /FS_MedSAM2/sam2/build_fsmedsam2.py under /path/to/SAM2/sam2/build_fsmedsam2.py.
3. After changing the saved_npz_path = '/path/to/saved_npz' and ckpt_path = '/path/to/ckpt', eval FS_MedSAM2:

cd /path/to/SAM2/notebooks
python infer_fsmedsam2_by_slice.py # infer 1S1Q
python infer_fsmedsam2_by_volume.py # infer S1SFQ from top
python infer_fsmedsam2_by_volume_from_middle.py # infer S1SFQ from middle

Citation

If you find these projects useful, please consider citing:

@misc{bai2024fsmedsam2exploringpotentialsam2,
      title={FS-MedSAM2: Exploring the Potential of SAM2 for Few-Shot Medical Image Segmentation without Fine-tuning}, 
      author={Yunhao Bai and Qinji Yu and Boxiang Yun and Dakai Jin and Yingda Xia and Yan Wang},
      year={2024},
      eprint={2409.04298},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.04298}, 
}

Questions

If you have any questions, welcome contact me at '[email protected]'