📄 VPTTA

This is the official pytorch implementation of our CVPR 2024 paper "Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation".

Environment

CUDA 10.1
Python 3.7.0
Pytorch 1.8.0
CuDNN 8.0.5

Our Anaconda environment is also available for download from Google Drive.

Upon decompression, please move czy_pytorch to your_root/anaconda3/envs/. Then the environment can be activated by conda activate czy_pytorch.

Data Preparation

The preprocessed data can be downloaded from Google Drive.

Pre-trained Models

The pre-trained models can be downloaded from Google Drive.

You can also train your own models from scratch following:

• OD/OC Segmentation

CUDA_VISIBLE_DEVICES=0 python OPTIC/train_source.py --Source_Dataset RIM_ONE_r3 --path_save_log OPTIC/logs --path_save_model OPTIC/models --dataset_root your_dataset_root

• Polyp Segmentation

Please refer to the Pytorch implementation of PraNet.

How to Run

Please first modify the root in VPTTA_OPTIC.sh and VPTTA_POLYP.sh, and then run the following command to reproduce the results.

# Reproduce the results on the OD/OC segmentation task
bash VPTTA_OPTIC.sh
# Reproduce the results on the polyp segmentation task
bash VPTTA_POLYP.sh

Citation ✏️

If this code is helpful for your research, please cite:

@article{chen2023vptta,
  title={Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation},
  author={Chen, Ziyang and Ye, Yiwen and Lu, Mengkang and Pan, Yongsheng and Xia, Yong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={},
  year={2024}
}

Acknowledgement

Parts of the code are based on the Pytorch implementations of DoCR, DLTTA, and DomainAdaptor.

Contact

Ziyang Chen ([email protected])