Here you will soon find the step-by-step instructions to install and work with the methodology presented in our paper.

To access the code for your own work, simply clone the repository and take what you need

git clone [email protected]:MedVisBonn/Segmentation-Distortion.git

The easiest way to replicate the results of the paper is via Docker. You can find all necessary template files in the docker directory. To build the image, copy the Dockerfile and build.sh and run

bash build.sh

Once the image is ready, modify the run.sh to mount datasets and run a container

bash run.sh

You should be attached to an interactive session within the container automatically.

The datasets we used in the paper are openly available at

• Calgary Campinas Brain MRI Dataset
• ACDC Challenge
• M&M Challenge

For both cardiac MRI datasets, we used the nnUNet pre-processing (branch nnunetv1) and the resulting batch generators.

The docker image is build on top of NVidia's PyTorch image 23.07 and thus needs NVidia drivers 530 or later. Additionally, to run GPU accelerated containers, the nvidia-container-toolkit has to be installed and configured. For more information, check out the official NVIDIA Container Toolkit documentation. Package dependencies are handled within the docker image.

To explore the project, you can run a jupyter lab server on the port specified in the Dockerfile.

jupyter lab --no-browser --allow-root --port 8888

Training and evaluation scripts are located in src/apps and a collection of examples can be found in src/demos, but this project is still work in progress. In case you are stuck, please don't hesitate to reach out.

This work is licensed under the GNU General Public License v3

@Inproceedings{Lennartz2023Segmentation,  
     year = {2023},  
     title = {Segmentation {Distortion}: Quantifying {Segmentation} {Uncertainty} {Under} {Domain} {Shift} via the {Effects} of {Anomalous} {Activations}},  
     type = {Inproceedings},  
     series = {LNCS},  
     volume = {14222},  
     publisher = {Springer},  
     booktitle = {Medical {Image} {Computing} and {Computer} {Assisted} {Intervention} ({MICCAI}), {Part} {III}},  
     doi = {10.1007/978-3-031-43898-1_31},  
     url = {https://link.springer.com/chapter/10.1007/978-3-031-43898-1_31},  
     author = {Jonathan Lennartz and Thomas Schultz},  
     pages = {316--325},  
}

For any questions or clarifications, feel free to reach out

lennartz (ät) cs.uni-bonn.de

(Please allow a couple of days for a response)