Dynamic PET Reconstruction on the GPU
Dynamic Positron Emission Tomography (PET) reconstructs the space-time concentration function of a radiotracer by observing the detector hits of gamma-photon pairs born during the radiotracer decay. The computation is based on the maximum likelihood principle, i.e. we look for the space-time function that maximizes the probability of the actual measurements. The number of finite elements representing the spatio-temporal concentration and the number of events detected by the tomograph may be higher than a billion, thus the reconstruction requires supercomputer performance. The enormous computational burden can be handled by graphics processors (GPU) if the algorithm is decomposed to parallel, independent threads, and the storage requirements are kept under control. This paper proposes a scalable dynamic reconstruction system where the algorithm is decomposed to phases where each phase is efficiently mapped onto the massively parallel architecture of the GPU.