2023 OHBM Abstract
This abstract, which serves as an introduction to my research improving statistical methods for fPET-FDG analysis, was accepted for a poster presentation at the 2023 Organization for Human Brain Mapping Conference in Montreal.
Bias of metabolic (de)activations introduced by polynomial detrending of fPET-FDG data
Sean Courseya,b, Grant A. Hartungb,c, Jonathan R. Polimenib,c,d, Jingyuan E. Chenb,c
a College of Science, Northeastern University, Boston, MA, USA;
b Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, USA;
c Department of Radiology, Harvard Medical School, Boston, MA, USA;
d Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, USA.
Introduction
Functional PET (fPET)-FDG is a new addition to the neuroimaging field that enables tracking dynamic changes of metabolic activity within a single scan (Hahn et al. 2016; Stiernman et al. 2021; Villien et al. 2014). Because FDG is administered through constant infusion and accumulates in cells over the course of the scan, the time activity curve (TAC) has a monotonically increasing baseline trend. To separate task-driven metabolic deviations from this baseline, previous studies have used various methods to remove the baseline, including detrending using the whole-brain mean TAC or polynomials. While polynomial detrending can help avoid removing meaningful global metabolic changes—an issue with using the mean TAC—polynomials can mischaracterize the shape of the TAC baseline such that the resulting regression leaves a consistently patterned residual time-course. Furthermore, when this residual time-course correlates with the task paradigm, it can introduce artifactual (de)activations that bias the statistical testing of the true task effect. Here, we combine simulations and analysis of empirical data to investigate such activation biases caused by baseline misspecification in fPET-FDG studies.
Methods
We first performed simulations to assess how the activation bias introduced by baseline misspecification varies with arterial input function (AIF), tissue biological parameters, the choice of baseline model, and task timing. Our simulation used a two-compartment model of the radiotracer kinetics (Villien et al. 2014). The biological ranges for the kinetic constants K1, k2, and k3 were taken from Heiss et al. 1984. The effect size of the activation bias was quantified for three different baseline models: whole-brain mean TAC and 2nd- and 3rd-order polynomials; and three different task paradigms associated with Regressors 1-3 shown in Fig. 1b. The fitting coefficient for the residual pattern was compared to that for the activation profile to calculate percent bias. Activation was modeled as a 30% increase in k3.
We used publicly available resting-state, i.e. task-free, fPET-FDG data from 25 subjects (Jamadar et al. 2020) to test whether mismodeled baseline trends can lead to spurious activations in real data. We applied the same three baseline models and task regressors from the simulations to this dataset and tested for significant metabolic activations at the group level.
Results
Using low-order polynomials to detrend resting-state data resulted in consistent residual patterns across subjects, and these residual patterns from the data agreed with simulations (Fig. 1a). Simulations indicated that the bias resulting from these patterns can be substantial but that the extent of this bias varies greatly depending on the task paradigm, polynomial order, kinetic constants, and AIF (Fig. 1b).
Fig. 2 shows the results of applying different task regressors to the resting-state data using mean-TAC and low-order polynomial baselines. The extent and spatial patterning of bias vary greatly, depending on the task paradigm and the polynomial order. For instance, Regressor 2 in conjunction with 2nd-order detrending produced significant artifactual activation in the right occipital lobe at the group level; and Regressor 3 in conjunction with 3rd-order detrending produced modest global artifactual activation. With a sufficiently large sample size, any bias introduced by baseline mischaracterization may potentially result in statistically significant activations at the group level (Smith et al. 2018).
Conclusion
Detrending fPET-FDG data using polynomials can leave residual patterns which correlate with task-regressors, biasing analysis results and potentially leading to artifactual activation. This bias depends on the task paradigm, polynomial order, and tracer kinetics—ranging from insignificant to producing artifactual activation with various spatial patterns. These results underscore the importance of choosing detrending methods carefully in fPET-FDG studies.
Figures
