/MediCLIP

Official implementation of "MediCLIP: Adapting CLIP for Few-shot Medical Image Anomaly Detection (MICCAI 2024 Early Accept)"

Primary LanguagePythonMIT LicenseMIT

MediCLIP

💡 This is the official implementation of the paper "MediCLIP: Adapting CLIP for Few-shot Medical Image Anomaly Detection"(MICCAI 2024) [arxiv].

MediCLIP is an efficient few-shot medical image anomaly detection method, demonstrating SOTA anomaly detection performance with very few normal medical images. MediCLIP integrates learnable prompts, adapters, and realistic medical image anomaly synthesis tasks.

🔧 Installation

To run experiments, first clone the repository and install requirements.txt.

$ git clone https://github.com/cnulab/MediCLIP.git
$ cd MediCLIP
$ pip install -r requirements.txt

Data preparation

Download the following datasets:

Unzip them to the data. Please refer to data/README.

🚀 Experiments

To train the MediCLIP on the BrainMRI dataset with the support set size is 16:

$ python  train.py --config_path config/brainmri.yaml  --k_shot 16

To test the MediCLIP on the BrainMRI dataset:

$ python  test.py --config_path config/brainmri.yaml  --checkpoint_path xxx.pkl

Replace xxx.pkl with your checkpoint path.

Code reference: [CLIP] [CoOp] [Many-Tasks-Make-Light-Work].

🔗 Citation

If this work is helpful to you, please cite it as:

@inproceedings{zhang2024mediclip,
      title={MediCLIP: Adapting CLIP for Few-shot Medical Image Anomaly Detection}, 
      author={Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou},
      year={2024},
      eprint={2405.11315},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}