ReportQA is a clinical-related and flexible radiology report evaluation framework. This repository provides the complete pipeline for:
- structured report extraction,
- QA generation,
- QA filtering,
- QA-based evaluation.
ReportQA supports radiology report datasets across any imaging modality, anatomical region, and language. We release processed datasets including CTRG-Brain (brain CT), CT-RATE (chest CT), AMOS-MM (abdominal CT), and MIMIC-CXR on Hugging Face.
Filtered QA pairs can be used not only for radiology report evaluation, but also as standalone benchmarks for evaluating vision-language models.
# 1. clone and navigate
git clone https://github.com/MSIIP/ReportQA.git
cd ReportQA
# 2. create a conda environment, activate it and install packages
conda create -n reportqa python=3.11
conda activate reportqa
pip install -r requirements.txtTake CTRG-Brain-zh as an example:
# generate qas from free-form reports
bash scripts/generate_qas/generate_qas.sh# deploy the judge model with vLLM
bash scripts/deploy_vllm/deploy_vllm.sh
# self-filter & report-based filter
bash scripts/filter/filter_ctrg_brain_zh.sh# install `ms-swift` first: https://github.com/modelscope/ms-swift
# zero-shot inference
bash scripts/infer/internvl/infer_ctrg_brain_zh.sh# evaluation & scoring
bash scripts/eval/eval_ctrg_brain_zh.sh@article{shi2026reportqa,
title={ReportQA: QA-Based Radiology Report Evaluation},
author={Shi, Yiming and Yang, Shaoshuai and Chen, Xi and Li, Haolin and Zhang, Hengyu and Jiang, Che and Wang, Kaiwen and Zhu, Xun and Xie, Dong and Wang, Fei and others},
journal={arXiv preprint arXiv:2606.15037},
year={2026}
}We would like to express our gratitude to the following resources:
- CTRG - Brain & Chest CT dataset with Chinese & English radiology reports.
- CT-RATE - Chest CT dataset with English radiology reports.
- AMOS-MM - Abdominal CT dataset with English radiology reports.
- MIMIC-CXR - Chest X-ray dataset with English radiology reports.
- RadEvalX - Chest X-ray dataset with English radiology reports.