PT  - JOURNAL ARTICLE
AU  - Heather Liu
AU  - Jenny Ceccarini
AU  - Bart De Laat
AU  - Johan Lataster
AU  - Inez Myin-Germeys
AU  - Evan Morris
TI  - A comparison of kinetic models for PET imaging of cortical dopamine release induced by a task.
DP  - 2018 May 01
TA  - Journal of Nuclear Medicine
PG  - 502--502
VI  - 59
IP  - supplement 1
4099  - http://jnm.snmjournals.org/content/59/supplement_1/502.short
4100  - http://jnm.snmjournals.org/content/59/supplement_1/502.full
SO  - J Nucl Med2018 May 01; 59
AB  - 502Objectives: 8F-fallypride PET studies in humans have been successfully analyzed with the linear simplified reference region model (LSSRM)1 to quantify dopamine (DA) release after behavioral challenges. LSSRM does not describe time-varying neurotransmitter kinetics explicitly, but does account for time-dependence in radiotracer efflux from tissue. Intuitively, it would seem that LSSRM might hold an advantage over traditional kinetic models for circumstances when DA varies with time during a scan. We sought to understand the sensitivity of LSSRM to DA changes captured in dynamic PET data and to compare the performance of LSSRM to traditional time-invariant kinetic models. We simulated ­­18F-fallypride uptake kinetics in the cortex in the presence of a sustained dopaminergic behavioral challenge and examined the ability of the models to describe the simulated data. Methods: We performed realistic noisy simulations of voxel-level 18F-fallypride PET data in the cortex in the presence of DA competition induced by a behavioral stress task. Simulations were performed using an extended compartmental model according to Morris et al.2. We simulated data for scans at baseline and stress conditions. We modeled stress-induced DA release as gamma-variate curves of different shape profiles (varying sharpness and rise time) beginning 100 min into a 166 min scan, based on Lataster et al.3 (Fig. 1). The mean peak concentration of DA release was set to double the baseline value, based on published results3. Tracer kinetic parameters were based on literature4. Population variability, based on a cohort of 12 subjects, was introduced into tracer kinetic constants and also applied to the magnitude of DA release. Within-subject variability was modeled according to: εi = scale ∙ sqrt(PETi/Δti). PETi is the signal at a single time point, i, without decay correction; Δti is the duration of the time frame; εi is the standard deviation of the additive error in the TAC which was scaled to resemble the variability in experimental 18F-fallypride data3. For all DA release shapes (Fig. 1), simulated data from the stress condition only was fitted with LSSRM to estimate the peak DA release, γ. We also fitted simulated data from both the baseline and the stress conditions with the multilinear reference tissue model (MRTM) to determine change in binding potential (ΔBP). We compared the t-scores (for γ and ΔBP) representing the effect sizes estimated by the two models for stress-induced DA release. To compare goodness of fit between LSSRM and MRTM, we compared the weighted sum of squared residuals (WSSR). For consistency, WSSR was calculated only from the fits of the stress condition data. Results: LSSRM and MRTM produced comparable t-scores over a wide range of sharpness, α, and rise time for DA signals, indicating that LSSRM may not provide an advantage over a traditional time-invariant models if the only goal is the detection of a DA phenomenon (Fig. 2, left). However, LSSRM consistently produced smaller WSSR over the entire range of DA signal shapes considered (Fig. 2, right). In terms of ability to describe DA curve shapes, the goodness of fit for MRTM degrades as rise time increases, while LSSRM remains consistent. Conclusions: For 18F-fallypride PET studies that incorporate a sustained behavioral stress challenge, simulations suggest that LSSRM may not be advantageous over traditional time-invariant models if detection of DA change is the goal. However, LSSRM consistently outperformed MRTM in terms of fitting accuracy. This may be relevant if the study goal is to characterize the PET signal and/or to deduce the DA signal shape or duration. Further, LSSRM and other time-varying models may be the safer choice when the duration of the DA response and/or its shape are pertinent or unknown. Our simulations highlight that t-score is not an indicator of goodness of fit but merely a measure of the magnitude of an estimated parameter relative to its variability.