Official repository for the IEEE TMI paper “Large Language Model with Region‑Guided Referring and Grounding for CT Report Generation”

git clone https://github.com/zhi-xuan-chen/Reg2RG.git
cd Reg2RG
conda create -n reg2rg python=3.9
conda activate reg2rg
pip install -r requirements.txt

The RadGenome‑ChestCT dataset used in this work can be downloaded from the Hugging Face hub: https://huggingface.co/datasets/RadGenome/RadGenome-ChestCT

• Base LLM: Llama‑2‑7b‑chat‑hf
• Pre‑trained & auxiliary checkpoints: Reg2RG

After the data and checkpoints are ready, you can inference the model on RadGenome‑ChestCT with our pre‑trained weights.

1. Create a custom configuration file [your_config_name].sh in configs/test_radgenome/ (you can copy and modify configs/test_radgenome/jhcpu7.sh).
2. Run:

   cd scripts
   bash test_radgenome.sh [your_config_name]

After inference, evaluate the generated reports with the scripts in the evaluation/ directory.

All scripts read file paths from variables inside the script.
Please open each file and set the path (e.g. results_path) to your own CSV results before running.

1. Strip region prefixes

   python evaluation/rm_region_text.py

2. Compute Natural Language Generation (NLG) metrics

   python evaluation/hf_nlg_evaluation.py

3. Compute Clinical Efficacy (CE) metrics

   python evaluation/ce_evaluator_ct2rep/ce_evaluation.py

1. Region prediction accuracy

   python evaluation/region_pred_acc.py

2. Split into region‑specific reports

   python evaluation/parse_report_to_region.py

3. Compute region‑level NLG metrics

   python evaluation/hf_nlg_evaluation_region.py

4. Compute region‑level CE metrics

   python evaluation/ce_evaluator_ct2rep/ce_evaluation_region.py

You can fine‑tune the model on your own dataset using our pre‑trained checkpoint.

1. Add your dataset class to src/Dataset/, following the style of the existing datasets.
2. Create a config [your_config_name].sh in configs/train_radgenome/ (copy and adjust configs/train_radgenome/jhcpu7.sh).
3. Run the training script:

   cd scripts
   bash train_radgenome.sh [your_config_name]

If you find our work useful, please consider citing it:

@article{chen2025reg2rg,
  title   = {Large Language Model with Region-Guided Referring and Grounding for CT Report Generation},
  author  = {Chen, Zhixuan and Bie, Yequan and Jin, Haibo and Chen, Hao},
  journal = {IEEE Transactions on Medical Imaging},
  year    = {2025},
  publisher = {IEEE}
}

This project builds on the RadFM repository and the LLaMA 2 model.
We sincerely thank the original authors for their invaluable contributions to the community.