Uncertainty quantification in non-rigid image registration via stochastic gradient Markov chain Monte Carlo

We provide the source code used in the research published in the MELBA Special Issue: UNSURE 2020.

Usage

Set-up

• NiBabel
• matplotlib
• numpy
• pandas
• PyTorch
• scikit-learn
• SimpleITK
• tvtk

Registration

To align images use the following command:

python run.py -vi 1 -mcmc 1 -d device_id -c config.json

config.json specifies the configuration to use for training, incl. the path to input images and the values of hyperparameters. The input images must have a .nii.gz extension and will be automatically resized to dimensions specified in the configuration file. The directory with the input images must contain subdirectories seg with the segmentations and masks with the image masks.

To resume registration:

python train.py -r path/to/last/checkpoint.pth

Citation

If you use this code, please cite our paper.

Daniel Grzech, Mohammad Farid Azampour, Huaqi Qiu, Ben Glocker, Bernhard Kainz, and Loïc Le Folgoc. Uncertainty quantification in non-rigid image registration via stochastic gradient Markov chain Monte Carlo. MELBA 2021, Special Issue: UNSURE 2020, 1–25.

@article{Grzech2021,
    author = {Grzech, Daniel and Azampour, Mohammad Farid and Qiu, Huaqi and Glocker, Ben and Kainz, Bernhard and {Le Folgoc}, Lo{\"{i}}c},
    title = {{Uncertainty quantification in non-rigid image registration via stochastic gradient Markov chain Monte Carlo}},
    year = {2021},
    journal = {MELBA},
    number = {Special Issue: UNSURE 2020},
    pages = {1--25}
}