Impact of depth-of-interaction and spatially-variant point-spread-function model on the performance of ultra-high...

To implement a spatially-variant point spread function (VarPSF) model for the NeuroEXPLORER (NX) brain positron emission tomography (PET) scanner in the open-source Yale Reconstruction Toolkit for PET (YRT-PET), and investigate the interplay between depth-of-interaction (DOI) resolution and resolution modeling. Approach. PSF parameters were estimated from simulated and physically measured point sources. Three primary DOI configurations were evaluated: eight-layer DOI (8DOI), two-layer DOI (2DOI), and no DOI (noDOI), with an additional 4DOI comparison in the off-center simulated Derenzo study. Reconstructions using VarPSF, a spatially-invariant PSF measured at the field-of-view (FOV) center (SinglePSF), and no PSF modeling (noPSF) were assessed using simulated Derenzo and brain phantoms, a physical Derenzo phantom, and one non-human primate (NHP) scan. Main results. In simulated and physical Derenzo studies, VarPSF yielded higher peak-to-valley ratios (PVRs) than SinglePSF and noPSF at the final evaluated iterations, with pronounced gains at large radial offsets. In the simulated brain phantom, VarPSF reduced the estimated transverse blur by 0.3 mm relative to SinglePSF and 0.5 mm relative to noPSF in 2DOI, and by 0.6 mm and 0.8 mm, respectively, in noDOI. In the NHP study, VarPSF improved contrast and structural delineation with reduced DOI encoding at approximately matched white-matter coefficient of variation, while its incremental benefit over SinglePSF was small with 8DOI at clinically relevant iteration settings. Significance. Spatially-variant PSF modeling can partially compensate for limited DOI information. The findings motivate combining modest DOI capability with accurate resolution modeling, while further validation is required for redesigned detectors and fully corrected acquisitions. The open-source implementation enables controlled investigation of the complementary roles of DOI encoding and resolution modeling.