This is the minimal reproduce code for the paper "Boosting Semi-supervised Image Segmentation with Global and Local Mutual Information Regularization" recently submitted to a journal.
We release the code, together with the well-preprocessed ACDC dataset for reviewers. The dataset should be keep private based on the dataset agreement and I will delete it once the reviewer process finishes.
Our code is based on deepclustering2 package, which is a personal research framework. It will automatically install all dependency on a conda virtual environment and without resorting to requirement.txt.
conda create -p ./venv python=3.7
conda activate ./venv
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch # install pytorch 1.6.0
pip install deepclustering2-2.0.0-py3-none-any.whl
python setup.py install
# all packages should be set properly automatically.In case of failure of running the experiments, please refer to requirement.txt to see the packages
cd semi_seg
# our proposed method
python main.py Data.labeled_data_ratio=0.05 Data.unlabeled_data_ratio=0.95 Trainer.num_batches=300 Trainer.max_epoch=100 Data.name=acdc Arch.num_classes=4 Optim.lr=0.0000001000 Trainer.name=udaiic Trainer.save_dir=udaiic/10.0_0.1 IICRegParameters.weight=0.1 UDARegCriterion.weight=10.0
# ps baseline (lower bound)
python main.py Data.labeled_data_ratio=0.05 Data.unlabeled_data_ratio=0.95 Trainer.num_batches=300 Trainer.max_epoch=100 Data.name=acdc Arch.num_classes=4 Optim.lr=0.0000001000 Trainer.name=partial Trainer.save_dir=ps
# fs baseline (upper bound)
python main.py Data.labeled_data_ratio=1.0 Data.unlabeled_data_ratio=0.0 Trainer.num_batches=300 Trainer.max_epoch=100 Data.name=acdc Arch.num_classes=4 Optim.lr=0.0000001000 Trainer.name=partial Trainer.save_dir=fs One can change the parameters on the cmd if needed.
Please refer to the default configuration in config/semi.yaml all set of controllable hyperparameters. All of them can be changed using cmd as above.
Based on different random seed, the ACDC performance varies within 1% in terms of DSC. Above scripts gives a DSC of ~85.5% for our proposed method vs 62.0% for ps and 89.2% for fs.