GPU-accelerated Monte Carlo based scatter correction in brain PET/MR

Abstract

A main advantage of PET is that it can provide truly quantitative results. In order to achieve this goal an accurate scatter correction is essential. Despite some drawbacks, the currently used single scatter simulation approach is clinically applicable, whereas Monte Carlo (MC) simulations have been shown to provide accurate results, but are still too slow to be used routinely. In this work, the high computing capabilities of graphical processing units (GPUs) are exploited to accelerate PET MC simulations in order to facilitate their use in clinical practice. Starting from voxelized images, the annihilation photons are tracked through to their detection in the simulated PET scanner geometry, while retaining the information on their associated scattering interactions. The scatter distribution provided by the simulation can subsequently be used to correct the measured datasets. The new GPU implementation was validated by comparison with GATE simulation results and found to provide equivalent accuracy. The acceleration factor on the GPU compared to current GATE simulations was 135 for a voxelized brain phantom study. The speedup of MC simulations provided by the graphics processors represents a major step towards a clinically feasible and physically accurate scatter correction.