XrayGPT is a state-of-the-art model for chest radiology report generation using large medical vision-language models. Built on top of BLIP-2 and MedCLIP, XrayGPT aligns a frozen visual encoder with a frozen large language model (LLM), Vicuna, using a linear projection layer. This repository extends XrayGPT for general-purpose medical report generation and Visual Question Answering (VQA).
Due to inconsistencies and incompatibilities among various libraries in the original codebase, a new environment is created to run the code in a Runpod container. The environment is based on Python 3.10, PyTorch 2.0.0, and CUDA 11.8.
Use the Runpod Template pytorch:2.1.0-py3.10-cuda11.8.0
and run the following commands to install the required libraries:
apt-get update -y && apt-get install zip unzip vim -y
python -m pip install --upgrade pip
pip install gdown
pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r hard_requirements.txt --no-deps
pip install pydantic==1.10.7
pip install hyperframe==5.2.0
pip install gradio==3.23.0
pip install safetensors==0.4.3
Below is a brief overview of the steps for fine-tuning the trained XrayGPT model. Instructions for training XrayGPT from scratch are provided in the original repository.
Publicly available datasets for medical report generation predominantly focus on chest X-ray reports, often derived from sources like MIMIC-CXR/OpenI. While these datasets are valuable, they lack diversity in terms of medical imaging modalities. To address this limitation and enhance the model's capabilities for multi-modality report generation and Visual Question Answering (VQA), we curated a unique dataset by integrating two distinct datasets: OpenI and ROCO.
OpenI Dataset: OpenI is a well-known resource provided by the Indiana University School of Medicine, comprising chest X-ray images paired with corresponding radiology reports.
- Kaggle Download: Link
- Description: Radiology reports and chest X-ray images
- Samples: 4,000
- Usage: Report generation (Chest X-ray)
ROCO Dataset: ROCO (Radiology Objects in COntext) is a multimodal medical image dataset enriched with descriptive captions, offering a broader spectrum of medical imaging scenarios.
- Description: Multimodal images with detailed descriptive captions
- Dataset Size: 8,000 samples (validation split used)
- Usage: Enables the model to generalize across various medical imaging modalities beyond chest X-rays.
By processing the OpenI and ROCO datasets using the OpenAI API and combining them, we created a comprehensive dataset suitable for training our model. The data integration resulted in a structured dataset stored in the dataset
folder, facilitating efficient training and evaluation processes.
- Scripts to create the dataset given in Data_Creation_Scripts folder
The final structure of the dataset folder is as follows:
dataset
├── image
| ├──1.jpg
| ├──2.jpg
| ├──3.jpg
| .....
├── filter_cap.json
Download the finetuned version of Vicuna-7B
from the original XrayGPT link. The final weights should be in a single folder with a structure similar to the following:
vicuna_weights
├── config.json
├── generation_config.json
├── pytorch_model.bin.index.json
├── pytorch_model-00001-of-00003.bin
...
Download the Minigpt-4 checkpoint from the trained XrayGPT model.
Here we fine-tuned a pretrained XrayGPT model on the dataset created above. The model was initially trained on the MIMIC and OpenI datasets in a two-stage training process.
Run the following command:
python3 train.py --cfg-path train_configs/xraygpt_openi_finetune.yaml
Download the pretrained XrayGPT checkpoints from the link and add this checkpoint in eval_configs/xraygpt_eval.yaml
.
Run the following command to launch the demo:
python demo.py --cfg-path eval_configs/xraygpt_eval.yaml --gpu-id 0
If you use this work, please cite the following original XrayGPT paper:
@article{Omkar2023XrayGPT,
title={XrayGPT: Chest Radiographs Summarization using Large Medical Vision-Language Models},
author={Omkar Thawkar, Abdelrahman Shaker, Sahal Shaji Mullappilly, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, Jorma Laaksonen and Fahad Shahbaz Khan},
journal={arXiv: 2306.07971},
year={2023}
}