Codes for the paper: "3D Cross-Pseudo Supervision (3D-CPS): A semi-supervised nnU-Net architecture for abdominal organ segmentation" by Yongzhi Huang*, Hanwen Zhang*, Yan Yan and Haseeb Hassan.

Fast and Low-resource semi-supervised Abdominal oRgan sEgmentation in CT (FLARE 2022)

Our framework is based on nnU-Net, so we strongly recommend you have a look at the repository of nnU-Net before starting with our code.

nnU-Net expects datasets in a structured format like the data structure of the Medical Segmentation Decthlon. The first step is to build dataset.json using dataset_convErsion.TaskXXX_TASKNAME.py.

We follow the network structure and other hyper-parameter settings automatically generated by nnU-Net, so the default ExperimentPlanner is enough.

python nnUNet_plan_and_preprocess.py -t TASK_ID --ssl

If you want to change any properties related to models or training hyper-parameters, you can inherit and override Class ExperimentPlanner3D_v21/ExperimentPlanner2D_v21, and modify the Class Trainer (In our work, it's nnUNetTrainerV2_SSL) accordingly. For more details, please refer to this guide in the nnU-Net repository.

python nnUNet_plan_and_preprocess.py -t TASK_ID -p YOUR_EXP_PLANNER --ssl

• 2D version:

  python run_training.py 2d nnUNetTrainerV2_SSL TASK_ID FOLD

• 3D version:

  python run_training.py 3d_fullres nnUNetTrainerV2_SSL TASKID FOLD

• 2D version

  python predict_simple.py -i INPUT_DIR -o OUTPUT_DIR -f FOLD -t TASK_ID -m 3d_fullres -tr nnUNetTrainerV2_SSL -p nnUNetPlansv2.1 -chk model_best

• 3D version:

  python predict_simple.py -i INPUT_DIR -o OUTPUT_DIR -f FOLD -t TASK_ID -m 2d -tr nnUNetTrainerV2_SSL -p nnUNetPlansv2.1 -chk model_best

• [CVPR 2021] Semi-Supervised Semantic Segmentation with Cross Pseudo Supervision

• nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation