PET image reconstruction using LOR-OSEM with a 3D spatially variant system matrix

A point source was used to sample the point spread function (PSF) at over 6000 locations within the field of view (FOV) of a General Electric Discovery ST PET scanner (DST) in 2D high sensitivity acquisition mode. These measurements were used to optimize the derivation of a system matrix for the DST. We found for 2D acquisition mode that a system matrix using a PSF with radial, depth, axial, and azimuthal dependence produced reconstructed images with greatly improved spatial resolution and contrast-noise ratios over the entire FOV, as compared to the use of a geometrically derived system matrix. The main improvements in resolution and contrast-noise resulted from the inclusion of depth dependence in the model, which accounts for large variations in sensitivity in the DST that are due to the septa and a span of 11 present in 2D acquisition. Finally, we determined that exploitation of symmetries, particularly along the axial dimension, allow a system matrix of similar quality to what we achieved with over 6000 samples to be created with just over 1000 samples, i.e. with an almost 80% reduction in sample size.