Point cloud dosimetry framework for preclinical microbeam radiation therapy

Abstract

Background:

Microbeam Radiation Therapy (MRT) is an emerging radiotherapy technique which is currently at the research stage. In order to further progress from research, toward clinical deployment, it is essential to develop a reliable and accurate dose engine such as Monte Carlo (MC) simulations.

Purpose:

MC execution times are far too long for practical, clinical applications. In previous studies, we used a 3D U-Net, trained with Geant4 MC simulations, to calculate the dose in digital rat phantoms. This choice of model imposes significant scalability challenges in the case of larger geometries. Casting MRT dose prediction as a 3D point cloud regression problem is a flexible and extensible solution to overcoming many of these hurdles. This problem formulation requires the use of point-based models which are unproven for the task of MRT dose prediction.

Methods:

In order to assay the viability of this family of models on this task, the SphereFormer is trained to accurately replicate the gold standard MC dosimetry on uniform voxel grids.

Results:

Furthermore, the benefits of the scalability of this method are demonstrated by the utilisation of out-of-field information to significantly improve over the existing state-of-the-art results on valley dose prediction, being accurate to within 3% for at least 84.1% of voxels compared with 78.2% for the baseline.

Conclusion:

This paper serves as a proof-of-concept study for the application of 3D point cloud methods to MRT dose prediction and marks the first time such a method have been applied to dosimetry in general.