Semi-supervised Contrastive Learning for Label-efficient Medical Image Segmentation

This is the pytorch implementation of paper "Semi-supervised Contrastive Learning for Label-efficient Medical Image Segmentation".

Setup

Environment

python=3.7.10
torch==1.8.1
torchvision=0.9.1
batchgenerator=0.21

Dataset

We will take the Hippocampus dataset as the example to illustrate how to do the preprocessing. Put the images .nii.gz files in ./data/Hippocampus/imgs folder and labels files in ./data/Hippocampus/labels.

cd dataset/prepare_dataset
python preprcocessing.py
python create_splits.py

After which,the images and labels will be restored together in .npy file with shape normalized to target size.

Run the codes

To run the global pretraining. Thie part is mainly based on the repository PyTorch SimCLR: A Simple Framework for Contrastive Learning of Visual Representations

bash run_simclr.sh

To run the supervised pixel-wise contrastive learning,

bash run_coseg.sh

To combine the above two pretraining, run run_simclr.sh first and the pretrained model will be saved at save/simclr/Hippocampus/ and set --pretrained_model_path ${the saved model path} in run_coseg.sh.

As for the segmentation finetuning, remember to load the saved model, and

bash run_seg.sh

Notice that in all the above three files, there is a parameter named train_sample, which means the percentage of labeled data to use.

If you use our codes or find our codes useful, please cite

@inproceedings{hu2021semi,
  title={Semi-supervised Contrastive Learning for Label-Efficient Medical Image Segmentation},
  author={Hu, Xinrong and Zeng, Dewen and Xu, Xiaowei and Shi, Yiyu},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={481--490},
  year={2021},
  organization={Springer}
}