Efficient voxel navigation for proton therapy dose calculation in TOPAS and Geant4

Abstract

A key task within all Monte Carlo particle transport codes is ‘navigation’, the calculation to determine at each particle step what volume the particle may be leaving and what volume the particle may be entering. Navigation should be optimized to the specific geometry at hand. For patient dose calculation, this geometry generally involves voxelized computed tomography (CT) data. We investigated the efficiency of navigation algorithms on currently available voxel geometry parameterizations in the Monte Carlo simulation package Geant4: G4VPVParameterisation, G4VNestedParameterisation and G4PhantomParameterisation, the last with and without boundary skipping, a method where neighboring voxels with the same Hounsfield unit are combined into one larger voxel. A fourth parameterization approach (MGHParameterization), developed in-house before the latter two parameterizations became available in Geant4, was also included in this study. All simulations were performed using TOPAS, a tool for particle simulations layered on top of Geant4. Runtime comparisons were made on three distinct patient CT data sets: a head and neck, a liver and a prostate patient. We included an additional version of these three patients where all voxels, including the air voxels outside of the patient, were uniformly set to water in the runtime study. The G4VPVParameterisation offers two optimization options. One option has a 60–150 times slower simulation speed. The other is compatible in speed but requires 15–19 times more memory compared to the other parameterizations. We found the average CPU time used for the simulation relative to G4VNestedParameterisation to be 1.014 for G4PhantomParameterisation without boundary skipping and 1.015 for MGHParameterization. The average runtime ratio for G4PhantomParameterisation with and without boundary skipping for our heterogeneous data was equal to 0.97: 1. The calculated dose distributions agreed with the reference distribution for all but the G4PhantomParameterisation with boundary skipping for the head and neck patient. The maximum memory usage ranged from 0.8 to 1.8 GB depending on the CT volume independent of parameterizations, except for the 15–19 times greater memory usage with the G4VPVParameterisation when using the option with a higher simulation speed. The G4VNestedParameterisation was selected as the preferred choice for the patient geometries and treatment plans studied.