Monte Carlo simulations play a crucial role for in-vivo treatment monitoring based on PET and prompt gamma imaging in proton and carbon-ion therapies. The accuracy of the nuclear fragmentation models implemented in these codes might affect the quality of the treatment verification. In this paper, we investigate the nuclear models implemented in GATE/Geant4 and FLUKA by comparing the angular and energy distributions of secondary particles exiting a homogeneous target of PMMA. Comparison results were restricted to fragmentation of (16)O and (12)C. Despite the very simple target and set-up, substantial discrepancies were observed between the two codes. For instance, the number of high energy (>1 MeV) prompt gammas exiting the target was about twice as large with GATE/Geant4 than with FLUKA both for proton and carbon ion beams. Such differences were not observed for the predicted annihilation photon production yields, for which ratios of 1.09 and 1.20 were obtained between GATE and FLUKA for the proton beam and the carbon ion beam, respectively. For neutrons and protons, discrepancies from 14% (exiting protons-carbon ion beam) to 57% (exiting neutrons-proton beam) have been identified in production yields as well as in the energy spectra for neutrons.
Online dose monitoring in proton therapy is currently being investigated with prompt-gamma (PG) devices. PG emission was shown to be correlated with dose deposition. This relationship is mostly unknown under real conditions. We propose a machine learning approach based on simulations to create optimized treatment-specific classifiers that detect discrepancies between planned and delivered dose. Simulations were performed with the Monte-Carlo platform Gate/Geant4 for a spot-scanning proton therapy treatment and a PG camera prototype currently under investigation. The method first builds a learning set of perturbed situations corresponding to a range of patient translation. This set is then used to train a combined classifier using distal falloff and registered correlation measures. Classifier performances were evaluated using receiver operating characteristic curves and maximum associated specificity and sensitivity. A leave-one-out study showed that it is possible to detect discrepancies of 5 mm with specificity and sensitivity of 85% whereas using only distal falloff decreases the sensitivity down to 77% on the same data set. The proposed method could help to evaluate performance and to optimize the design of PG monitoring devices. It is generic: other learning sets of deviations, other measures and other types of classifiers could be studied to potentially reach better performance. At the moment, the main limitation lies in the computation time needed to perform the simulations.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.