kakaobrain / CheXGPT

CheX-GPT

This is an official inference code of "CheX-GPT: Harnessing Large Language Models for Enhanced Chest X-ray Report Labeling" [arxiv]

Environment setup

We have experimented the implementation on the following enviornment.

• Python 3.11
• CUDA 12.1

conda create -n chexgpt python=3.11
conda activate chexgpt
pip install -r requirements.txt

Prepare dataset

TBU - a subset of MIMIC test data (500 reports and paired labels)

Model checkpoint

Download the model checkpoint from here [link] and place the model in the 'checkpoint' directory.

Command line

• Test
  • CE metrics are displayed

    python main.py mode=test

• Predict
  • Labeler outputs are saved in jsonline format

    python main.py mode=predict predict.output_path=${save_path}

• Inference
  • You can directly input CXR reports and check the labeler output.

    python inference.py

Citation

@article{gu2024chex,
  title={CheX-GPT: Harnessing Large Language Models for Enhanced Chest X-ray Report Labeling},
  author={Gu, Jawook and Cho, Han-Cheol and Kim, Jiho and You, Kihyun and Hong, Eun Kyoung and Roh, Byungseok},
  journal={arXiv preprint arXiv:2401.11505},
  year={2024}
}

License

• The source code of CheX-GPT is licensed under CC-BY-NC 4.0 License.
• The CheX-GPT model initialization utilized the ChexBert. For information regarding the licensing of ChexBert, please refer to the following links:
  • ChexBert License
  • Stanford Office of Technology Licensing

Contact

• Jawook Gu, [e-mail]
• Kihyun You, [e-mail]