Assessment of a white matter reference region for 11 C-UCB-J PET quantification

Introduction

Synaptic vesicle proteins, including the synaptic vesicle glycoprotein 2A (SV2A), are of interest for their roles in neurological and neurodegenerative disorders, including epilepsy ¹ and Alzheimer’s disease (AD). ² As such, SV2A has become a target of interest for development of positron emission tomography (PET) radioligands, in order to evaluate SV2A distribution in vivo. The current work employed ¹¹C-UCB-J; other SV2A-specific radioligands include ¹¹C-UCB-A, which has good affinity for SV2A but exhibits slow binding kinetics in pig ³ and monkey; ^{4 18}F-UCB-H, a molecule which has a relatively low affinity for SV2A (in vitro K_i = ∼ 50 nM) ⁵ and a low gray-to-white matter contrast in the nonhuman primate and human brain; and ¹⁸F-SDM-8, ⁶ the most-recently developed SV2A PET radioligand which has a high affinity (in vitro K_i = 0.58 nM), high gray-to-white matter contrast and has been used in nonhuman primates.

¹¹C-UCB-J has comparable SV2A affinity to ¹⁸F-SDM-8 and has been used for imaging in both nonhuman primates ⁴ and humans to investigate changes in synaptic density in epilepsy patients ⁷ and patients with mild cognitive impairment and AD. ⁸ Initial studies included dynamic PET acquisitions analyzed with compartment models using the metabolite-corrected arterial input functions. ⁹ However, in order for PET radioligands to be conveniently applied in the clinic setting, it would be preferred to apply noninvasive methods to attain meaningful quantitative measurements of the target of interest, such as reference region methods. Reference region methods allow for the quantification of the binding potential with respect to the nondisplaceable compartment, BP_ND. If an ideal reference region exists for a radioligand, then the total uptake in the reference region (V_T^Ref) should be equivalent to the nondisplaceable (i.e. free + non-specifically bound radioligand) uptake in the remaining regions of interest (ROIs), V_ND. In other words, there should be no specific or displaceable uptake in the reference region.

For noninvasive quantification with ¹¹C-UCB-J, the white matter region of the centrum semiovale (CS) was proposed as a reference region, due to its low PET uptake compared to cortical and subcortical gray matter in nonhuman primates and humans and its low concentration of SV2A compared to gray matter regions of the baboon brain shown by in vitro methods. ⁷ However, as reported in the first-in-human study, during displacement scans using SV2A specific drug levetiracetam (LEV), ¹¹C-UCB-J concentration decreased by about 10–15% in the CS, indicating the presence of some degree of specific binding in that region. ⁷ In this work, we further carefully characterize ¹¹C-UCB-J binding in the CS, initially by optimizing the CS ROI definition to minimize the inclusion of gray matter and any partial volume effects that would cause spill over and thus an overestimation of baseline V_T. Additionally, convergence using the ordered subset expectation maximization (OS-EM) reconstruction algorithm was investigated by increasing the iterations used to reconstruct the PET image; this would minimize any positive bias that may be observed in the low-intensity white matter, surrounded by the high-intensity gray matter. Lastly, baseline and blocking ¹¹C-UCB-J scans were used to estimate the V_ND of the gray matter, and the validity of using CS V_T to estimate V_ND in calculating BP_ND was evaluated.

Materials and methods

Image quantification and analysis

Regional time activity curves were analyzed using the one-tissue compartment model (1TCM),^9,15 using the arterial input function and the first 60 min of time activity data. The total volume of distribution V_T, the sum of the specific volume of distribution V_S, and the nondisplaceable volume of distribution V_ND, was estimated as K₁/k₂ for each ROI and scan condition for the following 23 gray matter (GM) ROIs: amygdala, anterior cingulum, anterior insula, caudate, cerebellum, dorsolateral prefrontal cortex, frontal cortex, fusiform, hippocampus, insula, occipital cortex, orbitofrontal cortex, pallidum, parahippocampal region, parietal cortex, posterior cingulum, posterior insula, putamen, substantia nigra, temporal cortex, thalamus, ventral striatum, ventromedial prefrontal cortex, and the multiple CS ROIs. Blood volume fraction was not included in the kinetic evaluation, as it has been shown to have negligible effects on both K₁ and V_T as shown in Supplemental Figure 1 of Finnema et al. ⁹

For each blocking study, the receptor occupancy (r) and the V_ND were determined using the Lassen Occupancy plot. ¹⁶ In summary, the difference in V_T between a blocking scan and the corresponding baseline scan were plotted against the baseline V_T values for each GM ROI. Graphically, the occupancy is defined as the slope of the best-fit line, while the x-intercept is equal to the V_ND. Mathematically, if V_T^Base= V_ND + V_S and V_T^Drug= V_ND + (1 − r) × V_S, then the difference between these expressions is V_T^Base − V_T^Drug= r × V_S. Substituting V_T^Base− V_ND for V_S

V T Base - V T Drug = r × ( V T Base - V ND )

(1)

This analysis was performed for each pair of baseline and blocking scans, using reconstructed images with 2, 3, and 4 OS-EM iterations (all with 30 subsets).

To examine possible heterogeneities in regional V_ND and V_S, parametric images of these parameters were created using parametric V_T images of baseline and blocking scans, which were generated using voxel-wise fitting of the 1TCM as described previously, ⁹ and the plasma concentration of BRV during blocking ¹¹C-UCB-J scans. Parametric V_T images were produced by voxel-wise fitting of 1–60 min of dynamic PET images to the 1TCM and were registered to MNI template space. Images from nine pairs of baseline and acute-post-dose scans with BRV, reconstructed with 2 OS-EM iterations, were used to estimate V_ND and V_S at each voxel i using V_T data from each subject j, with equation (2)

V T , i , j = V ND , i + ( 1 - r j ) × V S , i

(2)

where r_j = 0 for baseline scans, and r_j was calculated using the average drug concentration in blood during the post-dose scan, c, and the IC₅₀ for BRV, which was fixed at 0.46 µg/mL, ¹⁰ as shown in equation (3)

r j = c j c j + IC 50

(3)

From this, parametric V_ND and V_S images were created with voxel-by-voxel linear regression of the 18 V_T values (9 subjects × 2 scans each) to estimate V_ND and V_S for each voxel.

The PET outcome measure of binding potential, BP, is indicative of in vitro measurements B_max/K_D, or the product of target density and the affinity of the radioligand for the target. BP_ND is the binding potential with respect to the nondisplaceable compartment, and can be calculated by using the following

BP ND = V T - V ND V ND

(4)

If a reference region exists such that the total uptake in the reference region represents the free and nondisplaceable uptake in the gray matter, then equation (4) becomes

BP ∧ ND = V T - V T Ref V T Ref

(5)

To determine the efficacy of substituting the V_T of the CS for V_ND, BP_ND was calculated from baseline scans for the 23 ROIs for a subset of subjects (N = 7) using the extrapolated V_ND from the Lassen plots for each subject ( BP ND True ), as well as the V_T in the CS at 2 and 4 iterations (2i and 4i) of 30 subsets ( BP ND 2 i ∧ and BP ND 4 i ∧ ). Calculated BP_ND values were plotted against true BP_ND to assess discrepancies from ground truth by deviation from the unity line.

For evaluating changes in outcome measures due to increasing OS-EM iterations, one-way ANOVA (or its non-parametric equivalent, Friedman test) tests were used with post-hoc testing for multiple comparisons when appropriate. Differences between two values (i.e., baseline CS V_T vs. V_ND) were analyzed using the Student’s t test or non-parametric equivalent, Mann Whitney test. For subjects who received chronic dosing of BRV with a single baseline scan and multiple post-dose scans, the baseline CS V_T values were used only once in estimation of group statistics, and the blocking CS V_T values and V_ND estimates from these studies were averaged before evaluating group statistics.

Results

CS V_T at higher OS-EM iterations

Representative baseline and blocking V_T images are shown in Figure 2(a) and (b). Occupancy measures derived from the Lassen plot, shown for this representative subject in Figure 2(c), range from 50.0% to 87.3% and do not significantly change with higher OS-EM iterations. Increasing the number of OS-EM iterations had negligible effects on baseline and blocking V_T values in gray matter ROIs, with percent increases of 1.7 (±0.8)% and 1.3 (±0.3)% from 2 to 4 iterations. This was not the case in white matter, where increasing the number of OS-EM iterations decreased the CS baseline V_T by 8.5 ± 5.8%, from an average (±SD) of 4.58 ± 0.64 mL/cm³ at 2i to 4.21 ± 0.74 mL/cm³ at 4i (N = 13, p < 0.0001, Friedman/ANOVA, Figure 3(a)). Additional OS-EM iterations did not have a significant effect on CS V_T during blocking scans, changing from 3.60 ± 0.54 at 2i to 3.64 ±0.65 mL/cm³ at 4i (N = 28, p = 0.1292, Friedman/ANOVA, Figure 3(b)). OPEN IN VIEWER

Figure 2.

(a) Baseline and (b) blocking parametric V_T images from a representative ¹¹C-UCB-J scan. (c) Lassen Occupancy plots using V_T values from two (blue circles), three (red squares), and four (green triangles) iterations of the representative subject.

OPEN IN VIEWER

Figure 3.

(a) Individual (N = 13) centrum semiovale (CS) volume of distribution, V_T, at baseline from images reconstructed with 2, 3, and 4 OS-EM iterations of 30 subsets. Results from post-hoc testing with Dunn’s multiple comparison test are shown as p values. (b) Individual (N = 28) CS V_T under blocking conditions from images reconstructed with 2, 3, and 4 OS-EM iterations. (c) Average baseline (filled circles) and blocking (open circles) CS V_T at 2 OS-EM iterations. (d) Average baseline (filled circles) and blocking (open circles) CS V_T at 4 OS-EM iterations. For (c) and (d), results from Mann-Whitney U test are shown as p values.

Discussion

¹¹C-UCB-J is a PET ligand that binds selectively to SV2A and has been proposed as a quantitative imaging biomarker for synaptic density. The CS has been proposed as a reference region in ¹¹C-UCB-J PET quantification due to its minimal concentration of SV2A compared to gray matter regions in the baboon brain; however, displaceable activity was observed in the CS during LEV displacement studies in humans. ⁷ In this study, we examined the utility of CS as a reference region. First, the size and location of the CS ROI were optimized and the number of OS-EM iterations used to reconstruct the PET images was increased. These changes decreased, but did not eliminate the displaceable ¹¹C-UCB-J uptake in the CS, with the V_T in the CS overestimating GM V_ND by 36%. Thus, if the CS is used as a reference region, BP_ND is negatively biased by 40–55% (2 iterations) or 35–40% (4 iterations). However, the CS region may still be the best general-purpose reference region available for ¹¹C-UCB-J PET image quantification, given the correlations of CS V_T with gray matter V_ND values.

This work also presents a method to create parametric maps of V_ND and V_S using V_T maps from baseline and blocking scans, plasma drug concentrations, assuming a constant IC₅₀ value for BRV. A potential limitation is the use of a fixed IC₅₀ value across all brain regions. However, when extending the two-parameter model (equation (2)) to fit regional V_T data from these nine pairs of scans to a three-parameter model to also estimate regional IC₅₀ in addition to V_ND and V_S, the GM regions of the frontal cortex, temporal cortex, hippocampus, and putamen provided highly consistent IC₅₀ estimates: average (±SD) of 0.38 (±0.03) µg/mL. Reliable estimates of IC₅₀ in WM could not be obtained.

We hypothesize that the difference between white matter uptake and the true nondisplaceable ¹¹C-UCB-J uptake in gray matter is attributable to three causes: (1) lack of convergence with OS-EM algorithm in white matter, (2) a small amount of true specific binding in the white matter, and (3) lipophilicity differences between gray and white matter.

OS-EM is an iterative reconstruction algorithm which is used routinely in PET imaging due to its lower noise compared to filtered backprojection algorithms.^12,17 It has been reported in reconstructions of objects in which a low-radioactivity region is surrounded by a high-radioactivity region with high contrast, OS-EM tends to overestimate the activity within the low activity region. This bias is corrected, as the activity in the OS-EM image converges with increased iterations. ¹⁸ A caveat is that image noise increases with higher OS-EM iterations, so extra iterations should only be used for regions that might have not converged with standard methods. In the case of baseline ¹¹C-UCB-J scans, the CS has relatively low radioligand uptake compared to the high-radioactivity GM regions that surround it (Figure 2(a)). When we increase the number of iterations used to reconstruct these images, the positive CS bias with typical reconstruction parameters (2 iterations of 30 OS-EM subsets) was minimized. Increasing the number of iterations decreased the difference between baseline and blocking CS V_T, as well as between baseline CS V_T and GM V_ND; however, there is still a significant difference between these values after correcting for the overestimation bias caused by OS-EM convergence.

Although negligible concentrations of SV2A were reported using in vitro analysis in the CS in the baboon brain, ⁷ we reported displaceable in vivo ¹¹C-UCB-J uptake in human brain, indicating possible specific binding to SV2A in the CS. The slight specific binding to SV2A in the white matter may be due to the presence of SV2A protein in the white matter caused by the transport of synaptic vesicles or their precursors along the axon.^19,20 Additionally, SV2A has been reported to localize in mitochondrial membranes, albeit at a much smaller degree than in synaptic plasma membranes.^21,22 Though the concentration of SV2A due to these factors is likely minor, they may be contributing to all or part of the specific binding in the white matter, which is about 8% of the V_S in the gray matter. The displacement observed in the white matter could potentially be caused by the interaction of our tracer with a target other than SV2A; however, if this is the case, it is important to recognize that such sites were blocked by both antiepileptic drugs LEV and BRV, which are highly selective to SV2A.^23,24

White and GM are fundamentally different tissues within the brain. GM consists of neuronal cell bodies and glial cells, while white matter is predominantly composed of myelinated axons that project from a neuronal cell body to make connections to other neurons. The myelin that surrounds the axons in white matter contributes to an overall greater lipid content of this tissue compared to GM. ²⁵ Using our voxel-wise methods for V_ND estimation (Figure 5(d)), we found ∼ 20% higher nondisplaceable uptake in the white matter than in gray matter regions. This supports the hypothesis that, even without specific binding, the radioligand may behave differently between the two tissue types. However, as long as these differences in tissue characteristics are consistent within and across groups, the CS region may be a valuable reference region. The subjects in this study were recruited as healthy volunteers and were evaluated using a standard screening protocol. Clinical cases of neuroinflammation, neurodegeneration, white matter lesions, or vascular burden should be studied further to evaluate the use of a white matter reference region, in particular our CS_AAL ROI, in these populations of subjects.

Our group has shown significant differences in ¹¹C-UCB-J V_T and BP_ND, estimated from compartment modeling and arterial input function or using SRTM2 with the CS as a reference region, in the hippocampus in AD patients compared to cognitively normal individuals. ⁸ In that study, the V_T values in the CS region were very similar between groups. Assuming that GM V_ND is the same in both cohorts, we can calculate the BP_ND^True using the reported V_T values and the regional V_ND estimates derived from the parametric images reported here. As expected, the overall BP_ND estimates increased from 1.54 to 3.01 in the hippocampus of cognitively normal subjects, and from 0.87 to 1.89 in the hippocampus of AD patients. However, the between-group BP_ND differences were similar, from 43.5% different using CS V_T, to 37.1% different using the regional hippocampus V_ND derived from the parametric V_ND maps.

Other considerations while using the CS as a reference region include the sensitivity of our measures to the partial volume effect and to inaccuracies in scatter correction. The CS ROI used in this study was small (2 mL volume) and defined to ensure no GM was included in the ROI, as well as to minimize partial volume effect by avoiding white matter voxels that lay along the gray-white matter boundary in the CS. To assess whether this CS region was subject to any partial volume effects, partial volume correction (PVC) was applied to dynamic ¹¹C-UCB-J PET images using the iterative Yang’s method. ²⁶ This method produces a voxel-based correction by first creating a piece-wise uniform image with regional mean values, which is then convolved with the point-spread function of the system. The ratio between these images (before and after convolution) is applied to the original PET image, and the process iterates until convergence. In a subset of subjects (N = 42) used for CS ROI definition, this PVC had a minimal effect on CS values, with the baseline CS_AAL V_T decreasing by 2.4%, from 4.32 ± 0.67 mL/cm³ before PVC, to 4.22 ± 0.63 mL/cm³ after PVC. Thus, we believe the optimization of our white matter ROI minimized contributions of the partial volume effect. A caveat is that the PET data in this study were acquired on a Siemens HRRT, with image resolution of approximately 3 mm full width at half maximum. For data collected using a PET scanner that has poorer resolution, partial volume effects will likely contribute a noticeable bias in CS V_T, and thus PVC should be investigated. Note that if PVC is performed, it may be appropriate to use a larger CS ROI, to improve statistical reliability.

Scatter correction is an important aspect of PET quantification, with inaccurate scatter correction producing primarily an additive error to the images; this can produce a large percent error in low-activity regions. ²⁷ For baseline ¹¹C-UCB-J scans, with high gray:white contrast of ∼ 4:1, scatter correction accuracy is particularly important for CS quantification. Also, inaccurate scatter correction can produce different errors in baseline and blocking scans, due to their different contrast levels. To assess the sensitivity of our results to the accuracy of the single scatter simulation (SSS) correction method, we changed the scatter parameters in the reconstruction to assess the effect on baseline CS V_T estimates. Changing the lower level discriminator in SSS from 400 keV (default) to 375 keV led to a reduction in estimated scatter fraction and an increase in a representative subject’s baseline CS V_T from 4.95 to 6.11 mL/cm³. Thus, these measurements can potentially be sensitive to scatter correction and should be investigated further. As for many PET studies, scatter due to radioactivity sources that are out of the scanner field of view (FOV) may also cause inaccuracies in quantification, and corrections for this have been proposed. However, we believe that out of field (OOF) scatter will contribute minimally to image quantitation for these brain images on the HRRT, as the degree of OOF scatter is a function of the amount of radioactivity and the location of the source relative to the scanner FOV.^28,29 In the case of ¹¹C-UCB-J imaging, brain activity is very high, and based on whole body images of this tracer, there is no area in the upper body near the HRRT gantry with radioactivity levels approaching that of the brain, and thus will likely not introduce significant OOF scatter artifacts. In addition, due to the large axial FOV of the HRRT, there is typically ∼5 cm below the brain that is included in the scan. The lower axial slices in the images are incorporated in the scatter computation; OOF scatter would likely have the most effects on this region of the image, and since this is non-brain, we expect that OOF scatter will not introduce substantial bias to the brain measurements.

The midbrain region of the pons could also be considered as a potential reference region for SV2A-PET imaging with ¹¹C-UCB-J. However, in previous in vitro studies including Western blot, our group reported evidence of slight concentrations of SV2A in the pons. Additionally, we find that in nine acute blocking studies with LEV (N = 4) or BRV (N = 5), the pons has a higher average baseline V_T of 6.5 (0.4) mL/cm³ and has considerable blocking upon drug administration (48 ± 6%), about twice the amount in CS (22 ± 6%). Additionally, on average, the baseline V_T of the pons was 164.9 (±70.1)% greater than the GM V_ND. Thus, we believe the CS is likely to be a more valuable reference region for ¹¹C-UCB-J PET quantification.

Recent work by Koole et al. ¹⁵ also investigated CS as a reference region for ¹¹C-UCB-J quantification. In a study of 10 healthy subjects, baseline V_T values in the CS (denoted V_SO) were reported as 5.3 ± 0.6 mL/cm³, which is higher than the value reported here of 4.58 ± 0.74 mL/cm³. This is expected due to the poorer resolution of the PET system used in the former study compared to the HRRT, as the resolution likely contributes to greater partial volume effects and thus a higher spill-in from high radioactivity gray matter into the white matter. At average occupancy levels of 82.1 ± 14.5%, the average V_SO decreased to 4.9 ± 0.9 mL/cm³ during blocking scans, though this decrease of about 10% did not reach statistical significance. In our current study, we report a significant decrease between baseline and blocking V_T values in the CS, with average percent differences around 15%. When considering factors such as OS-EM convergence issues, partial volume effects, and differences in sample size and range of occupancies, our results and those of Koole et al. show considerable agreement.

Although there remains a significant difference between the CS and the V_ND in the gray matter of ¹¹C-UCB-J, given the correlation between these two estimates, the CS is likely a useful reference region for SV2A. Since SV2A is expressed in synaptic vesicles at virtually all synapses in the brain, there is no gray matter region available that would be devoid of the protein to allow for a true reference region to be used. Thus, the white matter, with specific uptake on the scale of less than 10% that of the gray matter, may be a valuable resource for estimating the nondisplaceable activity. Even though the CS uptake is predominantly due to nondisplaceable binding, it remains important to assess any differences in white matter uptake when investigating differences across disease groups, especially in diseases with white matter pathology.