Deformable Image Registration uncertainty quantification using deep learning for dose accumulation in adaptive proton therapy

Andreas Smolders, Antony lomax, Damien Charles Weber, Francesca Albertini

Keywords: Deformable image registration, Proton Therapy, Adaptive Planning, Uncertainty, Deep Learning

TL;DR: Presentation of an unsupervised deep learning method to predict the uncertainty associated with a given DIR solution.

Abstract: Deformable image registration (DIR) is a key element in adaptive radiotherapy (AR) to include anatomical modifications in the adaptive planning. In AR, daily 3D images are acquired and DIR can be used for structure propagation and to deform the daily dose to a reference anatomy. Quantifying the uncertainty associated with DIR is essential. Here, a probabilistic unsupervised deep learning method is presented to predict the variance of a given deformable vector field (DVF). It is shown that the proposed method can predict the uncertainty associated with various conventional DIR algorithms for breathing deformation in the lung. In addition, we show that the uncertainty prediction is accurate also for DIR algorithms not used during the training. Finally, we demonstrate how the resulting DVFs can be used to estimate the dosimetric uncertainty arising from dose deformation.

Dataset Code:

Data:

* The training data can unfortunately not be shared, as the policy at the Centre for Proton therapy does not allow to make patient images publicly available.

* The DIRLAB data is publicly available at https://www.dir-lab.com/

* The Non small cell lung cancer dataset was not acquired at the Centre for Proton therapy. We have requested the original clinic to make the data publicly available, but our request was denied.

Source code:

The source code is part of a larger framework to develop machine learning models. We do not want to make the framework fully publicly available for the following reasons:

* Parts of the code were not written by the authors of this manuscript
* A large part of the codebase is not used in this manuscript
* The code can anyhow not be run without the data

Therefore, we opted to cherry pick the most important files, which should contain all the implementation details. These can be found here: https://github.com/AndreasSmolders/WBIR2022. We are willing to provide larger parts of the code base upon request. Please contact the corresponding author in that case.