Background
Activated microglia, the resident immune cells in the brain, can be imaged by positron emission tomography (PET) by targeting mitochondrial 18-KDa translocator protein (TSPO) expression [
1‐
3]. [
11C]-PK11195 has been the most extensively used TSPO tracer; however, its low specific to non-specific binding ratio has led to the development of second-generation TSPO tracers [
4]. One such tracer is [
18F]FEPPA, which has been used to image neuroinflammation in a number of neurological diseases including Alzheimer’s disease [
1]. As with other TSPO tracers, [
18F]FEPPA quantification can be challenging due to the lack of a suitable reference region since microglia activation can occur throughout the brain. Consequently, the accepted method for quantifying [
18F]FEPPA uptake is to apply a two-tissue compartment model (2TCM), which requires generating a metabolite-corrected arterial input function (AIF) [
2]. In addition to invasive arterial blood sampling, 2–3 h of PET imaging is recommended to estimate the total distribution volume (
VT) with acceptable precession [
2].
The simultaneous estimation method (SIME) is an alternative approach that has the advantage of not requiring a reference region to quantify brain uptake. The principle of SIME is to analyze multiple tissue activity curves (TACs) simultaneously to estimate model parameters common to all regions. Initially developed as a modeling approach for estimating the AIF [
5], it was subsequently proposed for estimating a common non-displaceable distribution volume (
VND) across brain regions [
6,
7]. Assuming a common
VND has the advantage of reducing the number of independent parameters defining the TAC in each brain region. Furthermore, the estimated binding potential relative to a non-displaceable compartment (BP
ND) is not susceptible to scaling errors in the AIF since it is calculated from the ratio of distribution volumes, namely
\({\text{BP}}_{{{\text{ND}}}} = \left( {V_{{\text{T}}} - V_{{{\text{ND}}}} } \right)/V_{{{\text{ND}}}}\) [
6,
8]. This advantage is well suited to applications involving either population-based or image-derived input functions, which are prone to scaling errors. Indeed, an initial application to TACs and metabolite-corrected AIFs for [
11C]PBR28, a second-generation TSPO tracer, showed the ability of SIME to distinguish between high- and mixed-affinity binders (HABs and MABs, respectively) [
7]. In addition, if the input function is measured, the specific distribution volume (
VS) can be calculated from regional
VT and global
VND estimates:
VS =
VT −
VND. In certain patient populations, differences in
VND compared to controls were found, indicating that
VS may be a more sensitive indicator of TSPO activity than
VT [
8,
9].
Given the potential advantages of SIME, the overall objective of this study was to develop a minimally invasive SIME approach for [
18F]FEPPA PET imaging. As the first aim (“
Accuracy of SIME” section), retrospective data from healthy individuals were used to compare
VND and
VT estimates derived from the 2TCM and SIME to assess the accuracy of the latter [
2]. This analysis was conducted separately for HABs and MABs given the expected differences in TSPO binding. The second aim was to investigate if the scanning duration could be reduced from the recommended 2–3 h [
2] to 90 min without compromising the precision of BP
ND estimates obtained with SIME (“
Error Analysis” section). Next, experiments were conducted in a porcine model, in which arterial and venous blood samples could be drawn concurrently, to investigate if arterial sampling could be avoided by using venous samples for metabolite correction (“
Venous vs. Arterial Metabolite Correction” section). Finally, a feasibility study was conducted, by applying the minimally invasive SIME to data from healthy participants (“
Feasibility Study” section). The acquisition time was 90 min and image-derived input functions (IDIFs) were acquired instead of AIFs [
10]. Serial venous blood samples were used to scale each IDIF and for metabolite correction. The accuracy of the method was investigated by comparing group-wise regional BP
ND and
Vs estimates to those from the retrospective data.
Data acquisition
Participants were scanned for 90 min on a fully integrated PET/MRI system (Biograph mMR, Siemens Healthcare GmbH, Erlangen, Germany) using a 16-channel (12 head, 4 neck array) PET-compatible RF coil. Following the intravenous administration of [
18F]FEPPA (209.20 ± 52 MBq), 90 min of list-mode PET data was acquired. Dynamic images were reconstructed into 52 time frames (1 × 20 s, 12 × 5 s, 8 × 15 s, 4 × 30 s, 5 × 60 s, 10 × 120 s, 11 × 300 s, and 1 × 279 s) with a matrix size of 344 × 344 × 127 and voxel size of 1.043 × 1.043 × 2.032 using the Siemens e7 tools and an iterative reconstruction algorithm without point-spread function modeling (ordered subset expectation maximization (OSEM) with 3 iterations, 21 subsets and 2 mm Gaussian post-smoothing filter). The PET data were corrected for decay, scatter, and dead time, while attenuation correction was performed using the RESOLUTE method applied to ultrashort echo-time MR images [
13]. The imaging protocol included the acquisition of
T1-weighted structural images using a magnetization-prepared rapid gradient-echo (MP-RAGE) sequence (matrix size: 256 × 256 × 240, voxel size: 0.8 × 0.8 × 0.8 mm, echo time: 2.25 ms, repetition time: 2400 ms, and flip angle: 8°). Venous blood samples for metabolite analysis were manually drawn at the antecubital fossa (ACF). Catheters were inserted left and right ACF in both forearms, one for injection, the other for blood draw. Venous blood samples were collected at 2.5, 7, 12, 20, 30, 45, 60, and 90 min.
For PET quantification using SIME, each participant’s IDIF was extracted using a semi-automated software algorithm that extracts a subject-specific mask of the carotid arteries from high-resolution anatomical MR images, which were subsequently registered to corresponding PET images [
10]. The whole-blood IDIF extracted from the carotid mask was corrected for partial volume errors and spill-in contamination [
14,
15] and scaled to whole-blood venous samples acquired at 45, 60, and 90 min. Venous blood samples collected were analyzed to calculate BPRs and remaining [
18F]FEPPA fractions according to a previously published method [
2]. Following the procedure used in the animal validation experiments, radioactivity in whole-blood and plasma samples was measured using a well counter and the unmetabolized [
18F]FEPPA fraction was measured by chromatography. A biexponential function was used to determine the BPRs and a Hill function was used to estimate the unmetabolized [
18F]FEPPA fraction. These measurements were calculated for each subject individually. A metabolite-corrected plasma IDIF was generated by applying the correction functions for BRP and the remaining [
18F]FEPPA fraction to the whole-blood IDIF [
2].
Image preprocessing
Dynamic PET images were processed using SPM12 (
https://www.fil.ion.ucl.ac.uk/spm/). The images were realigned, registered to the anatomical image, skull stripped, and spatially normalized to Montreal Neurological Institute (MNI) space using SPM’s unified segmentation-based normalization approach [
16]. The six ROIs used for SIME and the 2TCM were defined based on the AAL atlas [
17] and extracted using an in-house MATLAB script.
Error analysis
Numerical simulations were conducted to examine the accuracy and precision of parameter estimates obtained by applying SIME to 90 min of [
18F]FEPPA data. The analysis was based on data used to determine the accuracy of SIME. First, the residual between the best fit of the 2TCM and a TAC was calculated at each time frame and normalized by the mean activity. The procedure was performed for all regions (i.e., for FL, TL, Ceb, Tha, Ins, and Cau) and across all 19 participants [
8,
18]. The average residual for each region and time frame was calculated across participants. Next, one participant’s dataset was selected, and the initial values of
K1,
k2,
k3, and
k4 for each of the six ROIs were defined by the best fit of the 2TCM. Simulations were performed by adding Gaussian noise to the theoretical TACs. The magnitude of the noise at each time frame was scaled to reflect the average residual calculated for that time frame and region. The SIME procedure was repeated 1000 times to generate histograms of best-fit estimates of
k2,
k3, and
k4 for each region and a common
VND, which were subsequently used to calculate
VT, BP
ND, and
VS. The precision of each parameter was defined by the coefficient of variation (COV) of the histogram and the accuracy by the agreement between the mean estimate and the input value.
Statistics
Analysis of variance (ANOVA) was conducted to compare HABs and MABs VT estimates determined by the 2TCM and SIME, with ROI as the within-subject variable and genotype as the between-subject variable. Similarly, an ANOVA was used to determine differences between VND estimates obtained by the two methods. Repeated-measures ANOVAs were used to compare regional BPND estimates for the two scan durations (90 and 120 min), as well as the corresponding VND estimates. Similarly, repeated-measures ANOVAs were used to assess the effects of reducing the number of ROIs on the VND and regional BPND estimates. Differences in inter-subject variability were assessed using an F test to compare variances. The analysis was performed for VT estimates from the 2TCM and SIME, as well as BPND and VND estimates from SIME for the two scan durations.
Pearson’s correlations analysis was conducted between venous and arterial blood-to-plasma ratios (BPR) and remaining [18F]FEPPA fractions. The difference in the regression slope from a value of one was performed using a paired t test. For the feasibility study, a two-way repeated-measures ANOVA was used to compare BPND and VS estimates from the group involving IDIFs to the group involving AIFs. Each analysis was conducted over six ROIs with binding affinity as the between-subject variable. All statistical tests involving multiple comparisons included Bonferroni corrections. Analysis was performed in SPSS (IBM, Armonk, NY, USA, version 27) using the F test to assess the homogeneity of variance. Statistical significance was assessed based on p < 0.05. All values are reported as mean ± standard deviation (SD).
Discussion
Noninvasive quantification of TSPO uptake in the brain by PET is challenging due to the lack of a suitable anatomical reference region. The necessity for arterial blood sampling to measure the AIF is particularly concerning when working with clinical populations with cognitive and behavioral challenges, such as patients with dementia and other neuropsychiatric disorders. This study investigated a minimally invasive SIME-based approach aimed at simplifying the steps required to quantify regional [
18F]FEPPA uptake. The main findings were that SIME enabled the imaging duration to be reduced to 90 min from the original 2–3 h required for full kinetic analysis [
2] and accurate estimates of regional BP
ND and
VS were generated by SIME using IDIFs that incorporated venous metabolite correction.
To evaluate the accuracy of SIME, it was applied to a [
18F]FEPPA dataset that included metabolite-corrected AIFs in order to compare
VND and
VT estimates to those obtained by standard kinetic modeling. All regional
VND values were within ± 22% of the overall average. Furthermore, there was no significant difference between
VND estimates from SIME and the 2TCM, suggesting that the assumption of a uniform
VND across brain regions is reasonable for [
18F]FEPPA. The accuracy of estimating
VND will depend on selecting an appropriate number of ROIs with a range of kinetic behaviors [
6,
19]. For this study, ROIs were selected based on literature values of the rate constants for [
18F]FEPPA [
2] and visible differences in their TACs. Permutation analysis demonstrated that
VND estimates with similar accuracy could be obtained for four, five, and six ROIs; however, significantly better precision was achieved with six.
Similar to the
VND results, good agreement was found between SIME and the 2TCM with respect to regional
VT. No significant differences between estimates from the two approaches were found in any of the six ROIs analyzed, and both approaches detected significantly larger
VT values for HABs compared to MABs. These results are in good agreement with Schain et al., 2018 who used SIME for quantification with the second-generation TSPO tracer [
11C]PBR28 [
7]. A further advantage of SIME is that by reducing the number of fitting parameters, the inter-subject variability was reduced by roughly 40% compared to the 2TCM. This result agrees with previous studies involving multiple tracers that reported greater reliability in parameter estimation with SIME [
6,
7,
20].
By taking advantage of the greater precision provided by SIME, the possibility of reducing the scan duration required for accurate [
18F]FEPPA quantification was investigated. The results presented in Fig.
2 demonstrate that reducing the scan time from 120 to 90 min had no significant effect on the accuracy of either regional BP
ND or brain-wide
VND. Furthermore, BP
ND exhibited the same dependency on genotype as
VT. The ability to estimate regional BP
ND precisely was confirmed by the Monte Carlo simulations (Fig.
3). These simulations indicated that for average residuals based on measured TACs, only small differences were found between the average estimates of BP
ND and the corresponding input values (average bias of 1 ± 0.3%). A potential limitation of this error analysis is that Gaussian noise does not account for possible covariance between TACs from separate ROIs. To assess potential covariance, the correlation between residuals from different TACs was calculated. The average correlation was weak to moderate (0.33 ± 0.42), indicating that the noise model was reasonable. An alternative approach would be to use empirical residuals in the simulations as proposed by Plavén-Sigray et al.; however, the number of datasets in the current study was not large enough to generate a sufficient number of simulations [
8].
The application of SIME, such as with [
11C]PBR28, has typically involved using a population-based AIF [
6,
7]. A disadvantage of this approach is the possibility of changes to the shape of the AIF related to binding affinity or disease [
9,
18]. In the current study, an alternative minimally invasive approach was investigated that involved extracting the IDIF for each participant and using serial venous blood samples for metabolite correction. Mabrouk et al. [
21] demonstrated that accurate estimates of regional
VT could be obtained using IDIFs scaled by an arterial blood sample. Based on these results, the current study used IDIFs scaled to late venous samples (i.e., 45, 60, and 90 min) when arterial and venous [
18F]FEPPA concentrations are expected to be in equilibrium. Serial venous samples were also used for metabolite correction, as confirmed in the animal study that showed good agreement between arterial and venous measures of BPR and remaining [
18F]FEPPA fraction (Fig.
4). The greatest error was found at the earliest sampling time, indicating that venous and arterial [
18F]FEPPA concentrations had yet to reach equilibrium [
20]. However, Fig.
4 illustrates that the magnitude of the error was relatively small (i.e., 10.8 ± 6.2% for the remaining [
18F]FEPPA fraction at 2.5 min). This was confirmed by comparing BP
ND estimates from the two groups of healthy participants. No significant difference in BP
ND was found in any of the ROIs and there was good agreement between the two groups in terms of the difference between HABs and MABs. The need for venous sampling does add complexity to the imaging procedure. A possible solution would be to incorporate a model of the metabolite pool into SIME, although this comes at the cost of increasing the dimensionality of the cost function [
22]. Note that these animal studies did not account for possible variations in the shape of the AIF due to TSPO polymorphisms; however, there is no evidence of polymorphisms in pigs [
23].
Although BP
ND has the advantage of being insensitive to scaling errors in the input function, recent studies have reported greater variability in BP
ND compared to
VS [
8]. Furthermore,
VND has been found to change in some patient populations, which would be a confounder for interpreting changes in
VT and suggests that
VS is a more appropriate marker of TSPO binding [
9]. By collecting venous blood samples in the current study, it was possible to scale individual IDIFs in order to estimate regional
VS. Similar to BP
ND,
VS was found to be sensitive to genotype and regional estimates obtained by the minimally invasive approach were in good agreement with those obtained from the dataset that included AIFs.
A limitation of this work was that the feasibility study did not involve measuring the AIF and consequently, validating the minimally invasive SIME was based on a comparison to the retrospective dataset. The good agreement in terms of regional BP
ND and
VS, as well as brain-wide
VND from the two datasets, indicates that the minimally invasive approach can measure these parameters accurately. Another consideration is the age difference between participants in the two datasets; the retrospective data consisted of participants in middle age, while the mean age of the participants recruited for the minimally invasive imaging protocol was above 70 y. Despite this age difference, normal aging is not associated with increased [
18F]FEPPA uptake [
24]. A final limitation was the lack of a patient population to investigate if the approach is sensitive to disease-specific changes in TSPO binding. The consistently greater values of
VT,
VS, and BP
ND in HABs compared to MABs demonstrate that the approach is sensitive to changes in specific binding; however, it would be prudent to investigate its accuracy in patient populations.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.