Semi-supervised-Cardiac-Image-Segmentation-via-Label-Propagation-and-Style-Transfer

Paper

This is the official implementation of the paper:

Semi-supervised-Cardiac-Image-Segmentation-via-Label-Propagation-and-Style-Transfer,

which won the 2nd place in MICCAI 2020 Multi-Centre, Multi-Vendor & Multi-Disease Cardiac Image Segmentation Challenge (M&Ms)

Result

Usage

Please change the input and output paths before running the code!!!

Prepare data
- Download the data from MICCAI 2020 M&Ms Challenge.
- Rename the data files. In Data_Proprocess, run python rename.py and python rename_unlabeleddata.py.
Propagate the label to unlabelled volumes by registration. In Data_Proprocess, run sh register.sh.
Train a preliminary model with both labelled and unlabelled data.
- Preprocess the data referring to the usage of nnUNet.
- In nnunet_semi_sup, train the model by python run/run_trainin.py 3d_fullres nnUNetTrainerV2 TASK_ID fold=all.
Transfer the style of data by histogram matching. In Data_Proprocess, run python match_to_target_histogram.py.
Finetune the model with with both original and trasferred data.
- Preprocess the data referring to the usage of nnUNet.
- Copy the preliminary model to the finetune model path.
- In nnunet_semi_sup, finetune the model by python run/run_trainin.py 3d_fullres nnUNetTrainerV2Finetune TASK_ID -f all -c.
inference on the test data by python inference/predict_simple.py -i INPUT_PATH -o OUTPUT_PATH -t TASK_ID -f all -tr nnUNetTrainerV2Finetune

Citation

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

@inproceedings{zhang2020semi,
  title={Semi-supervised Cardiac Image Segmentation via Label Propagation and Style Transfer},
  author={Zhang, Yao and Yang, Jiawei and Hou, Feng and Liu, Yang and Wang, Yixin and Tian, Jiang and Zhong, Cheng and Zhang, Yang and He, Zhiqiang},
  booktitle={International Workshop on Statistical Atlases and Computational Models of the Heart},
  pages={219--227},
  year={2020},
  organization={Springer}
}

Acknowledgement

The implementation is based on the out-of-box nnUNet.