This repository is the official implementation of Segmentation Consistency Training:Out-of-Distribution Generalization for Medical Image Segmentation. The paper is available here
To install requirements:
pip install -r requirements.txt
To train an instance of a predictor with a given model architecture and an ID number (to name the saved models), run:
python train.py [Model Architecture] [ID]
To train ten predictors for each model architecture, run:
bash run_paper_experiments.sh
To evaluate the models as trained using the above script, run:
python eval.py [ID lower] [ID upper]
This will evaluate the models given with IDs in the range [ID lower, ID ipper]
We demonstrate that Segmentation Inconsistency Training improves generaliation by a statistically significant margin (p>0.99) on all three tested out-of-distribution datasets.
##Contact information For questions or help, feel free to contact me at birk.s.torpmann-hagen@uit.no.