TumorCP: A Simple but Effective Object-Level Data Augmentation for Tumor Segmentation

Paper

This is the implementation for the paper:

TumorCP: A Simple but Effective Object-Level Data Augmentation for Tumor Segmentation

Accepted to MICCAI 2021

Usage

• Data Preparation

  • Download the data from MICCAI 2019 KiTS Challenge.

  • Convert the files' name by

  python dataset_conversion/Task040_KiTS.py

  • Preprocess the data by

  python experiment_planning/nnUNet_plan_and_preprocess.py -t 40 --verify_dataset_integrity

  • Extract the tumor region in advance by

  python extract_tumors.py

• Configuration

  The default configuration of TumorCP is in ./configuration.py. You can modify the parameters in the Trainer. Here are the examples: nnUNetTrainerV2_ObjCPAllInter and nnUNetTrainerV2_ImgDAObjCPAllInter.

• Train

  Train the model by

  python run/run_training.py 3d_fullres nnUNetTrainerV2_ImgDAObjCPAllInter 40 0

TumorCP is integrated with the out-of-box nnUNet. Please refer to it for more details.

• Test

• inference on the test data by

python inference/predict_simple.py -i INPUT_PATH -o OUTPUT_PATH -t 40 -f 0 -tr nnUNetTrainerV2_ImgDAObjCPAllInter

Citation

If you find this code and paper useful for your research, please kindly cite our paper.

@inproceedings{yang2021tumorcp,
  title={TumorCP: A Simple but Effective Object-Level Data Augmentation for Tumor Segmentation},
  author={Yang, Jiawei and Zhang, Yao and Liang, Yuan and Zhang, Yang and He, Lei and He, Zhiqiang},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={579--588},
  year={2021},
  organization={Springer}
}

Acknowledgement

TumorCP is integrated with the out-of-box nnUNet.