This repo holds the Pytorch implementation of PairAug:

[CVPR2024] PairAug: What Can Augmented Image-Text Pairs Do for Radiology?

• Create a new conda environment

conda create --name pairaug python=3.10
source activate pairaug

• Clone this repo

git clone https://github.com/YtongXie/PairAug.git
cd PairAug

• Install packages for image generation

pip install -r requirements_T2I.txt

• Download MIMIC-CXR-JPG dataset (Need to be a credentialed user for downloading)
  

• Put the image data under data/MIMIC_images_ori/ should be like:

    ├── p19_p19995997_s50123635_6fa953ea-79c237a5-4ca3be78-e3ae6427-c327e17b.png
    ├── p19_p19996061_s58482960_87923de8-5595ad44-eaa89d38-610e97e2-42cacf04.png
    ├── ...

• Run python data/reports_jsonl.py to generate the report data lists data/metadata_train.jsonl.
• Put the report data under data/MIMIC_reports_ori/ should be like:

    ├── p19_p19995997_s50123635_6fa953ea-79c237a5-4ca3be78-e3ae6427-c327e17b.txt
    ├── p19_p19996061_s58482960_87923de8-5595ad44-eaa89d38-610e97e2-42cacf04.txt
    ├── ...

• Download pretrained RoentGen weights and put it under pretrained/. Need to contact authors of paper RoentGen for download.
• Run

  python InterAug_Step1.py --start_index 0 --end_index 2200000

It aims to generate new patient reports via ChatGPT, and save them in data/InterAug/InterAug_reports.

• Run

  python InterAug_Step2_T2I.py --start_index 0 --end_index 2200000

It aims to generate inter-patient images via RoentGen model, and save them in data/InterAug/InterAug_images.

• Run

python IntraAug_Step1.py --start_index 0 --end_index 2200000

It aims to generate intra-patient reports via ChatGPT, and save them in data/IntraAug/IntraAug_reports.

• Run

python IntraAug_Step2_T2I.py --start_index 0 --end_index 2200000

It aims to generate intra-patient images based on generated and original reports, and save them in data/IntraAug/IntraAug_images.

• Install packages for data pruning

pip install -r requirements_MedClip.txt

• Run

python InterAug_Step3_fliter.py

It introduces a data pruning method w.r.t. the semantic alignment between generated InterAug image-report pairs to ensure the quality of the generated pairs.

• Run

python IntraAug_Step3_fliter.py

It introduces a hybrid consistency score between generated IntraAug image-report pairs to ensure the quality of the generated pairs.

Thanks to diffusers for the latent diffusion model, prompt-to-prompt for prompt-to-prompt stable diffusion model, RoentGen for pretrained weights, CheXzero for medical visual-language pre-training.

Yutong Xie (yutong.xie678@gmail.com)