A fast dynamic PET simulator for improved kinetic modeling estimation

Objectives Dynamic positron emission tomography (PET) plays an important role in in-vivo quantification of physiological processes in organs and tissues. In the field of oncology, the estimation of model based physiological parameters can provide a more accurate diagnosis, and aid in the screening, prediction, staging, treatment planning and segmentation of cancerous tumors, as well as enable better treatment follow-up. In this study we introduce a fast dynamic PET simulator called dPETSTEP that will allow researchers to better understand the bias and uncertainty tradeoffs as a function of the clinical environment and various post-processing choices, such as reconstruction parameters, post-filtering, and parameter fitting models. Useful examples are for educational purposes, to evaluate direct reconstruction algorithms, and optimal frame weighting for kinetic modeling.

Methods dPETSTEP is based on the simulation package PETSTEP, and is developed in Matlab. Using a parametric image as starting point, a pristine dynamic image is calculated based on the user-defined kinetic model, arterial input function and desired time sampling. Effects of the system’s point spread function, Poisson counting noise, uniform random counts, object dependent scatter counts and attenuation are all included to simulate a realistic dynamic PET scan from the dynamic pristine image. Fifteen replicates of a realistic head phantom with inserted tumor regions was simulated with both Monte Carlo (MC) and dPETSTEP. Furthermore, to evaluate dPETSTEP compared to simpler approaches, 15 simulation replicates of Gaussian noise was added to pristine time-activity curves (GAUSS). Voxel standard deviation (SD) vs. voxel value was investigated for both dPETSTEP and GAUSS, as well as difference maps relative the MC simulations. Finally, tumor ROI voxel value histograms for both uptake and parametric images from MC and dPETSTEP were compared.

Results The time taken to simulate 15 replicates of 28 frames of a 331x331x35-sized input parametric image of voxel size 1x1x4.25 mm was just under 47 min (2.9 GHz, 8 cores). Compared to the MC simulations, dPETSTEP was roughly 8000 times faster. The 4D image RMSE between the MC and dPETSTEP was 18% (0.17 kBq/ml) and 20% (0.20 kBq/ml) between MC and GAUSS. The normalized difference between the 4D dPETSTEP and MC images was on average 4%. Statistical analysis on the tumor ROI histograms showed no significant differences between the MC and dPETSTEP images for either uptake or parametric images (p<0.01). The scatter plot histograms of voxel SD vs. voxel value of dPETSTEP conformed quite well with MC, whereas GAUSS less so.

Conclusions We have demonstrated that dPETSTEP is able to simulate realistic 4D PET scans, where noise properties in both the dynamic images and subsequent parametric images conform very well with MC data, but in a fraction of the time, with results more accurate than a simple Gaussian approach. We believe dPETSTEP to be very useful for generating fast, simple and realistic results, however it uses a very simple scatter and random model that is not suitable for problems investigating these phenomena.