A Dempster-Shafer approach to trustworthy AI with application to fetal brain MRI segmentation

Trustworthy AI method based on Dempster-Shafer theory with application to fetal brain 3D T2w MRI segmentation.

System requirements

Hardware requirements

To run the automatic segmentation algorithms a NVIDIA GPU with at least 8GB of memory is required.

The code has been tested with the configuration:

• 12 Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz
• 1 NVIDIA GPU GeForce GTX 1070 with 8GB of memory

Operating system requirements

The code is supported on every operating system (OS) using docker. However, it has been tested only for

• Linux Ubuntu 18.04.6 LTS
• Linux Ubuntu 20.04.3 LTS

Installation

The installation is performed using docker:

First, install docker (see https://docs.docker.com/get-docker/).

Installation of the docker image

Install the docker image twai:latest using

sh build_docker.sh

This step takes a few minutes.

Create and start a docker container

Create a docker container for the docker image twai:latest that was previously built, using the command

docker run --ipc=host --gpus all -it -v <repository-path>:/workspace/trustworthy-ai-fetal-brain-segmentation -v <data-path>:/data --name twai twai:latest

where <repository-path> has to be replaced by the path of the git repository on your system and <data-path> has to be replaced by the path of a folder containing the data to be used for segmentation. This step creates a docker container called twai.

If you have already created the docker container twai, you can reuse it using the command lines

docker start twai
docker attach twai

The installation has been tested for

• Docker version 20.10.12, build e91ed57

How to use

Automatic Fetal Brain 3D MRI Segmentation

You can compute the automatic segmentations for the backbone AI, fallback, and trustworthy AI algorithms using the python script run_segment.py.

To learn more about the usage of the script, please see

python run_segment.py -h

We refer to the demo below for a detailed example.

Demonstration: example case

Fetal brain 3D MRI from a subset of the testing dataset can be downloaded at https://zenodo.org/record/6405632#.YkbWPCTMI5k

Put the folder \sub-feta001 of the first case in <data-path> on your system.

Start and attach the docker container (see above).

You can now compute the automatic segmentations for the backbone AI, fallback, and trustworthy AI algorithms using, inside the docker container

python run_segment.py --input '/data/sub-feta001/srr.nii.gz' --mask '/data/sub-feta001/mask.nii.gz' --ga 27.9 --condition 'Spina Bifida' --output_folder 'output/sub-feta001' --bfc

This step takes several minutes. You may need to adapt the paths depending where the folder \sub-feta001 is located inside <data-path>.

For more information about the argument of run_segment.py please run

python run_segment.py -h

The output can be found in the folder pointed by --output_folder (output/sub-feta001' in the example above). The output folder contains three main folders of interest:

• \backboneAI: output segmentation of the deep learning method nnU-Net
• \fallback: output segmentation of the atlas-based method
• \trustworthyAI: output segmentation of our trustworthy AI algorithm combining the output of the backbone AI and the fallback algorithms

Each of those folders should contain a unique segmentation file with the extension .nii.gz corresponding to the segmentation computed by the algorithm of the same name as the folder name.

Generating figures

The figures shown in the paper can be reproduced by running

sh run_make_all_figures.sh

After running this command, the figures will be in the folder \output.

How to cite

If you find this research useful for your work, please give this repo a star :star: and cite

• L. Fidon, M. Aertsen, F. Kofler, A. Bink, A. L. David, T. Deprest, D. Emam, F. Guffens, A. Jakab, G. Kasprian, P. Kienast, A. Melbourne, B. Menze, N. Mufti, I. Pogledic, D. Prayer, M. Stuempflen, E. Van Elslander, S. Ourselin, J. Deprest, T. Vercauteren. A Dempster-Shafer approach to trustworthy AI with application to fetal brain MRI segmentation

@article{fidon2022dempster,
  title={A Dempster-Shafer approach to trustworthy AI with application to fetal brain MRI segmentation},
  author={Fidon, Lucas and Aertsen, Michael and Kofler, Florian and Bink, Andrea and David, Anna L and Deprest, Thomas and Emam, Doaa and Guffens, Fr and Jakab, Andr{\'a}s and Kasprian, Gregor and others},
  journal={arXiv preprint arXiv:2204.02779},
  year={2022}
}