Monte Carlo (MC) simulations of X-ray image devices require splitting the simulation into two parts (i.e. the generation of x-rays and the actual imaging). The X-ray production remains unchanged for repeated imaging and can thus be stored in phase space (PhS) files and used for subsequent MC simulations. Especially for medical images these dedicated PhS files require a large amount of data storage, which is partly why Generative Adversarial Networks (GANs) were recently introduced. We enhanced the approach by a conditional GAN to model multiple energies using one network. This study compares the use of PhSs, GANs, and conditional GANs as photon source with measurements. An X-ray -based imaging system (i.e. ImagingRing) was modelled in this study. half-value layers (HVLs), focal spot, and Heel effect were measured for subsequent comparison. MC simulations were performed with GATE-RTion v1.0 considering the geometry and materials of the imaging system with vendor specific schematics. A traditional GAN model as well as the favourable conditional GAN was implemented for PhS generation. Results of the MC simulation were in agreement with the measurements regarding HVL, focal spot, and Heel effect. The conditional GAN performed best with a non-saturated loss function with R1 regularisation and gave similarly results as the traditional GAN approach. GANs proved to be superior to the PhS approach in terms of data storage and calculation overhead. Moreover, a conditional GAN enabled an energy interpolation to separate the network training process from the final required X-ray energies.
Keywords: Machine learning; Monte Carlo; Neuronal networks; X-ray.
Copyright © 2022. Published by Elsevier GmbH.