SLaVA-CXR: Small Language and Vision Assistant for Chest X-ray Report Automation [Paper]
conda create -n slava_cxr python=3.10 -y
conda activate slava_cxr
pip install --upgrade pip # enable PEP 660 support
pip install -e .
The SLaVA-CXR model can be downloaded in HuggingFace.
The training codes is made available. The training datasets are currently not available.
Evaluation dataset can be any chest X-ray frontal view image paired with a report.
We used MIMIC-CXR and IU-Xray datasets in our paper for the evaluation.
We have included IU-Xray questions for impression and findings section automation.
Please download IU-Xray dataset LINK.
CUDA_VISIBLE_DEVICES=0 python -m llava_phi.eval.model_vqa_slava_cxr \
--model-path ./SLaVA-CXR \
--question-file iuxray_sample_findings.jsonl \
--image-folder path_to_iuxray_images \
--answers-file findings_result.jsonl \
--conv-mode default \
--max_new_tokens 512
CUDA_VISIBLE_DEVICES=0 python -m llava_phi.eval.model_vqa_slava_cxr \
--model-path ./SLaVA-CXR \
--question-file iuxray_sample_impression.jsonl \
--image-folder path_to_iuxray_images \
--answers-file impression_result.jsonl \
--conv-mode default \
--max_new_tokens 256
@article{wu2024slava,
title={SLaVA-CXR: Small Language and Vision Assistant for Chest X-ray Report Automation},
author={Wu, Jinge and Kim, Yunsoo and Shi, Daqian and Cliffton, David and Liu, Fenglin and Wu, Honghan},
journal={arXiv preprint arXiv:2409.13321},
year={2024}
}
We used the LLaVA-Phi codes to train our model