Influence of Modeling Options on Kinetic Parameters for [18F]NOS PET Human Studies

Introduction: Tracer kinetic modeling is essential for accurate quantification of physiological parameters and is critically important for novel tracers. Replication across different sites is important for effective characterization of new agents. However, different sites use different modelling approaches, which may result in different conclusions. While much work has been performed to address scanner standardization, there has been little work to address the impact of options for modeling on PET quantitative outcomes. We therefore examined the impact of multiple modeling choices on neuroinflammation PET data from University of Pennsylvania with the novel tracer [¹⁸F]NOS [1] which binds to the inducible isoform of nitric oxide synthase that is induced by pro-inflammatory mediators. These studies included dynamic PET brain scans, MRI, and arterial blood samples with measurements of radiolabeled metabolites. To investigate this, we applied one-tissue and two tissue compartment (1TC, 2TC) modeling, and evaluated the following factors: region definition method, corrections for blood volume and time delay, data weighting in parameter estimation, and the fitting function for the plasma parent fraction.

Methods: Anonymized [¹⁸F]NOS DICOM image data and blood data from 5 healthy subjects were transmitted to the Yale PET Center database. Whole blood, plasma curves and HPLC data were encoded in Brain Imaging Data Structure (BIDS) format [2], which provides a systematic data structure for blood and imaging data. The PET scan duration was 60 min with frame lengths of 24x5, 6x10, 3x20, 2x30, 5x60 and 10x300s. Subject MRI images were first processed to extract a nonlinear registration [3] to register to the AAL template. Then a linear registration was used to register MRI to PET images. Time-activity curves (TACs) were then calculated for the following regions: cerebellum, frontal, occipital, parietal, and temporal lobes. Since the 2TC model was found to be unstable, the 1TC model, which generated good fits, was applied. The effects of changing the following factors were evaluated: 1) full AAL regions (unmasked) vs. regions masked by an SPM Grey Matter (GM) segmentation; 2) assuming 5% Blood Volume (BV) vs. 0% (BV=0); 3) correction for time delay between arterial and brain data (estimated by 1TC model fit to early portion of whole brain TAC) vs. no delay correction (Delay=0); 4) performing weighted fits (WT= Δt/(exp(λt)*C_T); WT=0 for first 15s) [4] vs. unweighted fits (NoWT); and 5) fraction of parent compound fitted with a sum of 2 exponentials or with an inverted gamma function (InvG) which can mimic a lag in metabolite production as with the Hill function. Comparison was made by assessing the % differences in K₁ and V_T to the values obtained by the standard modeling settings: 1) unmasked, 2) BV=5%, 3) estimated delay, 4) weighted fits, and 5) biexponential parent fraction fit.

Results: Mean parameter values (average across regions) using the standard analysis were K₁: 0.48 ± 0.09 mL/min/cm³ and V_T: 2.47 ± 0.28 mL/cm³. Table 1 shows the percent change in K₁ and V_T due to each modeling option. GM masking induced the largest and most consistent increases in K₁ (9.8%) and V_T (5.6%). Removing BV and delay corrections altered K₁ to a greater extent than V_T. In this case, data weighting had minimal effects, perhaps due to the robustness of the 1TC model. The effect of the choice of parent fraction model had the greatest intersubject variability.

Table 1. The mean percent difference across regions of K₁, V_T by adjusting 1 modeling factor at a time.

Conclusions: Factors influencing kinetic modeling results were studied for [¹⁸F]NOS, a tracer with rapid kinetics. In this case, changes in parameters were ~10% or less. For other tracers and models, especially tracers with slower kinetics or more complicated models, this type of analysis is important to optimize methodology.