Assessment of Gaussian and Multi-gaussian Kernel Models for the Reconstruction of TOF-PET Datasets with Event-by-Event Variable Time Resolution

Most current TOF-PET systems uses a global TOF uncertainty model during image reconstruction. Image artifacts can appear if the kernel model differs to much from the actual TOF kernel width, in particular when time resolution improves. Additionally, new detector designs such as hybrid Cherenkov/scintillation systems, detectors based on meta-materials in which different events can have a very different timing resolution, or 3 gamma imaging system bring several challenges as the TOF uncertainty model should be adapted in order to take advantage of the information carried by each detected event. This work focuses on (1) evaluating with Monte-Carlo simulated datasets a previously proposed analytical approach to estimate the SNR gain from a dataset with variable time resolution, and (2) study the performance of the use of different TOF kernel models for the reconstuction of datasets containing events with variable TOF information.Several datasets containing events ranging from 400 to 30 ps CRT were generated to produce 100 noise realizations. From these datasets, several mixes of labeled and unlabeled events were created, to generate datasets with either events with an unique CRT, or various mixes of events with different CRT. These datasets were then reconstructed using a gaussian TOF uncertainty model, as well as a weighed sum of gaussians. An image of pixel variances was computed for each approach from the set of 100 reconstructed images in order to derive SNR gain values.The evaluated analytical estimation approach was validated with these simulated results, both for datasets containing events with the same CRT or various mixes of different CRT. The use of a sum of gaussian provided a slightly better SNR when there is an acute difference in the proportion of events composing a mixed dataset, although this increase was not significant suggesting than using one gaussian model is still a reasonable approximation for these datasets.Future work will focus on the study of the different models with more realistic simulated datasets and ultra-fast time resolution.