A GPU-accelerated implementation of the MOLAR PET reconstruction package

MOLAR (Motion-compensation OSEM List-mode Algorithm for Resolution-recovery reconstruction) was written to provide the best possible images from ECAT HRRT PET data. Because of computational demands, MOLAR currently requires a computer cluster for practical use. Here we have applied GPU-acceleration via CUDA to all of the computationally intensive modules of the MOLAR package. Using an NVIDIA Tesla S1070-400 GPU system hosted by an HP xw8400 workstation, we evaluated the GPU-accelerated performance of the modules that perform boundary checking, forward and backprojection, photon scatter modeling and algorithm updates. We compared their performance to CPU-only versions of MOLAR for a range of total counts (500 k to 50 M). We found boundary checking to be up to 35 times faster using the GPU. Forward and backprojection ran 50 and 20 times faster, respectively, and scatter modeling was 200 times faster. Algorithm updates ran up to 15 times faster. The overall performance of the entire MOLAR package was approximately 40 times faster than the CPU-only code. These results show that MOLAR can be substantially accelerated using GPUs and can thereby be practically extended for use in high count and higher resolution applications, and for 4D parametric reconstructions.