Zhichao Sun, Yuliang Gu, Yepeng Liu, Zerui Zhang, Zhou Zhao, Yongchao Xu

We propose a Position-guided Prompt learning method for Anomaly Detection in chest X-rays (PPAD). PPAD leverages learnable text prompt and image prompt to minimize the gap between pre-training data and task-specific data. Through the position-guided prompts, the model can focus on various regions, simulating the diagnostic process of experts. Furthermore, we propose a Structure-preserving Anomaly Synthesis method (SAS) in the training phase.

• Installation
• Dataset
• Usage
• Model
• Results
• License
• Citation

Please follow the guide to install and set up the project.

# Clone the repository
git clone https://github.com/sunzc-sunny/PPAD.git

# Create conda environment
conda env create -f environment.yml
conda activate ppad

Please download the pre-training parameters of best_64_0.0001_original_35000_0.864.pt from CheXzero Google drive.

ZhangLab Chest X-ray
Please download the offical ZhangLab Chest X-ray benchmark from SQUID Google drive.

Stanford ChexPert
Please download the offical Stanford ChexPert benchmark from SQUID Google drive.

VinDr-CXR
Please download the offical Med-AD benchmark from DDAD Google Drive.
The Med-AD benchmark is organized using four public datasets, and VinDr-CXR is one of them.

Update the dataset path and pre-trained parameters with your specified paths.
Training the methods PPAD on datasets.

cd few_shot

python train_zhang.py      # Train PPAD on ZhangLab dataset
python train_chexpert.py  # Train PPAD on CheXpert dataset
python train_vincxr.py      # Train PPAD on VinDr dataset

Inference the method CheXzero on datasets.

cd zero_shot

python chexzero_zhang.py      # Test CheXzeor (baseline) on ZhangLab dataset
python chexzero_chexpert.py  # Test CheXzeor (baseline) on CheXpert dataset
python chexzero_vincxr.py      # Test CheXzeor (baseline) on VinDr dataset

Inference the method PPAD on datasets.

cd inference

python test_zhang.py       # Test PPAD on ZhangLab dataset
python test_chexpert.py   # Test PPAD on CheXpert dataset
python test_vincxr.py       # Test PPAD on VinDr dataset

PPAD on ZhangLab dataset Google Drive.
PPAD on CheXpert dataset Google Drive.
PPAD on VinDr-CXR dataset Google Drive.

*Methods marked with an asterisk (*) are reproduced in this work. Other results are either from their original paper or from SQUID. Mean and standard deviation are the results of five independent experiments.

*Methods marked with an asterisk (*) are reproduced in this work. Other results are either from their original paper or from DDAD. Mean and standard deviation are the results of five independent experiments.

This project is licensed under the MIT License

@misc{sun2024positionguided,
      title={Position-Guided Prompt Learning for Anomaly Detection in Chest X-Rays}, 
      author={Zhichao Sun and Yuliang Gu and Yepeng Liu and Zerui Zhang and Zhou Zhao and Yongchao Xu},
      year={2024},
      eprint={2405.11976},
      archivePrefix={arXiv}
}