Parallelization of MLEM algorithm for PET reconstruction based on GPUs

Abstract

The authors have implemented the Maximum Likelihood Expectation Maximization (MLEM) algorithm using Graphic Processing Units (GPU) and multiple cores (CPU) for the reconstruction of high spatial resolution positron emission tomography (PET) images. The massive amount of computation involved in precalculated system matrix limits the application of the MLEM reconstruction algorithm in practice, so the aim of this article is to simplify this process. The MLEM algorithm using Siddon and solid angle (SA) as projector was implemented in CPU and GPU-CUDA platform. Two implementations were developed, one computing the System Matrix on-the-fly and another with precalculated system matrix. Axial symmetries can reduce the amount of work in at least one order of magnitude. Parallelism then is exploited at the symmetry level, instead of at the lines of response (LOR) level. To further enhance the method performance the intersection between all rays in the LOR group and the image were precalculated. Later, a copy of this sub-image to local memory slice-by-slice, is carried out avoiding memory bottlenecks.