Adversarial Perturbations Improve Generalization of Confidence Prediction in Medical Image Segmentation
We introduce a straightforward adversarial training strategy that enhances the reliability of direct confidence prediction in medical image segmentation under realistic domain shifts.
- Clone the deployment branch of this repo (no code, only docker utils)
git clone --branch deploy --single-branch git@github.com:MedVisBonn/midl25.git- Build the image -
cd midl25/docker
bash build.sh- Create shared direotories for data and other files, adapt
docker/run.shaccordingly and create a container
bash run.sh- (optional) in the container, navigate to
/root/workplace/repos/midl25/and createresults/,pre-trained/monai-unets/andpre-trained/score-predictor/directories.
All applications can be run from bash files in src/apps.
- To train a U-Net, adapt
src/apps/train_unet.shand run it. - To train a score predictor, adapt
src/apps/trai_score_predictor.shand run it.
Further configurations can be found in src/configs/unet/monai_unet.yaml, src/configs/model/score_predictor.yaml and their respective trainer configs in src/configs/trainer/.
We evaluate our approach using two datasets: the SAML Dataset and the MNMS-2 Dataset. To work with these datasets, adapt the paths in the configuration files in src/configs/data to match your local environment. Any pre-processing is handled by the respective classes in src/dataset.
TBA
For questions, reach out to: lennartz (ät) cs.uni-bonn.de