This repository is a solution for the MICCAI FLARE2022 challenge. A detailed description of the method introduction, experiments and analysis of the results for this solution is presented in paper : Combining Self-Training and Hybrid Architecture for Semi-supervised Abdominal Organ Segmentation. As shown in the figure below, this pipeline consists of two parts: (a) pseudo-label generation for unlabeled data, which is implemented using PHTrans under the nn-UNet framework (for more information, see PHTrans); (b) a two-stage segmentation framework with Lightweight PHTrans. This repository is the code implementation of this part.
Download our repo and install packages:
git clone https://github.com/lseventeen/FLARE22-TwoStagePHTrans
cd FLARE22-TwoStagePHTrans
pip install -r requirements.txt
Download FLARE 2022 datasets. Generate pseudo-labels for unlabeled data based on the repository PHTrans. Modify the data path in the config.py file. Type this in the terminal to perform dataset processing:
python data_processing.py
Type this in terminal to run coarse segmentation train:
python coarse_train.py
Type this in terminal to run fine segmentation train:
python fine_train.py
Type this in terminal to Inference:
python predict.py -dp DATA_PATH -op SAVE_RESULTS_PATH