RadAdapt | ACL BioNLP 2023 (oral)

Official implementation of RadAdapt from Stanford University

• Title: RadAdapt: Radiology Report Summarization via Lightweight Domain Adaptation of Large Language Models
  
• Authors: Dave Van Veen*, Cara Van Uden*, Maayane Attias, Anuj Pareek, Christian Bluethgen, Malgorzata Polacin, Wah Chiu, Jean-Benoit Delbrouck, Juan Manuel Zambrano Chaves, Curtis P. Langlotz, Akshay S. Chaudhari, John Pauly
  
• Contact: {vanveen, cvanuden} [at] stanford [dot] edu
  

Environment

Use these commands to set up a conda environment:

conda env create -f env/environment.yml
conda activate radadapt

If your CUDA toolkit is older than 11.6 (display via nvcc --version), refer to env/README.md for modified instructions.

Usage

1. In src/constants.py, set your own project directory DIR_PROJECT.
2. Run a script, setting model and case_id as desired:
   • run_discrete.sh: generate output via discrete prompting.
   • train_peft.sh: fine-tune a model using a parameter-efficient method. LoRA is recommended.
   • run_peft.sh: generate output from a fine-tuned model.
   • calc_metrics.sh: calculate metrics on outputs.
3. To modify default parameters, create a new cases entry in src/constants.py.
4. To add your own dataset, follow the format in data/, which contains a subset of chest x-ray reports from Open-i.

Citation

@article{van2023radadapt,
  title={RadAdapt: Radiology Report Summarization via Lightweight Domain Adaptation of Large Language Models},
  author={Van Veen, Dave and Van Uden, Cara and Attias, Maayane and Pareek, Anuj and Bluethgen, Christian and Polacin, Malgorzata and Chiu, Wah and Delbrouck, Jean-Benoit and Chaves, Juan Manuel Zambrano and Langlotz, Curtis P and others},
  journal={arXiv preprint arXiv:2305.01146},
  year={2023}
}