Positron Emission Tomography (PET) is a medical functional imaging method providing information about the spatial distribution of biochemical or metabolic activity in the body. Before scanning, radioactive substance is injected into the subject of the measurement (e.g. human body, animals). Due to the beta plus decay of the injected substance, positrons are emitted, which, after meeting electrons, annihilate into gamma photon pairs.
The static PET reconstruction estimates the spatial function of the radiotracer density from the detected gamma photon pairs. The computation is based on the maximum likelihood principle. This means that we look for the spatial activity distribution function which most likely generates the same measurement that we are working on.
The reconstruction algorithm is an iterative approximation procedure. One iteration consists of simulating the measurement process (called forward projection), and improving the estimation (called back projection). The computation is highly complex, requiring the power of GPUs. This thesis work aims at efficient, accurate, fully-3D PET image reconstruction with the ML-EM algorithm on today’s GPUs. We examine the opportunities of improving both projections to achieve better image quality, while bearing the reconstruction time and memory requirements in mind.