Longitudinal Tau PET Using 18F-Flortaucipir: The Effect of Relative Cerebral Blood Flow on Quantitative and Semiquantitative Parameters

Visual Abstract

In vivo tau imaging allows for quantification of longitudinal changes in tau accumulation during the course of Alzheimer disease (AD) and can serve as a surrogate outcome measure in clinical trials. Several tau PET tracers are available for this purpose, of which 18 F-flortaucipir is the only one approved by the Food and Drug Administration (1)(2)(3)(4)(5). 18 F-flortaucipir PET images can be acquired using static or dynamic scanning protocols. Semiquantitative parameters such as SUV ratio (SUVr) can be derived from such a static PET scan. However, parameters derived from a dynamic PET scan, such as distribution volume ratio (DVR) or nondisplaceable binding potential (BP ND ), are fully quantitative and overall more accurate (6,7). Notwithstanding, dynamic protocols-because of the long scan duration-result in patient movement, lower patient comfort, and lower scanning efficiency. A compromise can be achieved by implementing a dual-time-window protocol in which overall scanning time is reduced by introducing a resting period during the scan while maintaining high quantitative accuracy (8)(9)(10).
SUVr has the advantage of practical applicability and relative computational simplicity (2)(3)(4)(5), while dynamic imaging studies provide more accurate measurements of specific binding and measure the relative tracer delivery (R 1 ), a proxy for relative cerebral blood flow (" 18 F-flortaucipir R 1 " section in the supplemental materials available at http://jnm.snmjournals.org) (7,(11)(12)(13)(14)(15). R 1 is important because blood flow changes can occur over time in AD because of disease progression or drug intervention. Longitudinal changes using SUVr may be biased by blood flow changes, whereas quantitative measures (BP ND ) are not (6,16). Currently, for 18 F-flortaucipir the sensitivity of SUVr for changes in blood flow has not been investigated. Therefore, with this study we compared SUVr and DVR/BP ND for 18 F-flortaucipir PET in a 2-y follow-up observational study. Second, we used simulations to investigate how larger changes in R 1 affect SUVr and DVR/BP ND .
The study protocol was approved by the Medical Ethics Review Committee of the Amsterdam UMC VU Medical center. All patients provided written informed consent before study participation.

Imaging
All subjects underwent 2 dynamic 18 F-flortaucipir PET scans, acquired on a Philips Ingenuity TF-64 PET/CT scanner, with a time period of 2.1 6 0.3 y (SCD) or 2.2 6 0.3 y (AD) between both scanning sessions. For SCD subjects, each scanning session consisted of 2 dynamic PET scans of 60 and 50 min, respectively, with a 20-min break in between (14,25). For AD patients, each scanning session consisted of 2 dynamic PET scans of 30 min and 20 min, respectively, with a 50-min break in between (9). BP ND , R 1 , and SUVr at 80-100 min were extracted in a prioridefined regions of interest (ROIs) in subject space using the Hammers and Svarer templates: Braak I/II (entorhinal), Braak III/IV (limbic), and Braak V/VI (neocortical). These ROIs align with neuropathologically defined regions (26) and are informative for tau PET in AD (27)(28)(29)(30).
For each parameter and ROI, we calculated percentage change using the following formula (DVR [BP ND 1 1] or SUVr associated with the follow-up and baseline scans, respectively): We repeated all analyses with partial-volume-corrected data using the iterative deconvolution method, as described previously (31)(32)(33).

Statistical Analyses
To allow for direct comparison with SUVrs, DVR was used for all analyses. Paired t tests were performed to assess differences between parameters and time points. Pearson correlation coefficients were computed to assess the correlation between percentage change in SUVr and DVR (all ROIs combined). Bland-Altman analyses were performed to assess bias and agreement between percentage change in SUVr and DVR (all ROIs combined). Analyses were performed in R software, version 4.0.2, and GraphPad Prism, version 9.1.0.
To explore whether the required sample size for (theoretic) future trials would differ when either quantitative or semiquantitative methods are used, sample sizes were calculated using GPower, version 3.1.9.7. For these analyses, we used a range of 0.5%-10% expected change in tracer retention over time, to inform on longitudinal study designs in the context of 18 F-flortaucipir. Sample sizes were calculated for SUVr and DVR, for all 3 ROIs (Braak I/II, III/IV, and V/VI). The differences between 2 dependent means (matched pairs) was calculated, with an a (error probability) of 0.05 and a power (1 2 b error probability) of 0.80. To adhere to the typical duration of clinical trials in AD, we calculated percentage change over an 18-mo period and used those SDs as input for the sample size calculations.

Simulations
Details on the methods used for simulations can be found in the Methods section of the supplemental materials. Table 1. In both AD and SCD, 18 F-flortaucipir SUVr were higher than DVR for all regions and at both time points (baseline and follow-up, all P , 0.001). Respective DVR, SUVr, and R 1 values are shown in Table 2 (SCD subjects and AD patients) and Supplemental Tables 2 and 3 (Ab-negative and -positive SCD subjects, respectively). The percentage overestimation of SUVr relative to DVR, for all regions and at both time points, is presented in Supplemental Table 4. Annualized percentage change in DVR and SUVr is presented in Supplemental Tables 5 and 6. No significant correlations between DVR or SUVr and R 1 were observed in either SCD or AD patients (Supplemental Fig. 1). Partial-volumecorrected data yielded essentially similar results; therefore, only noncorrected data will be presented further in the article.

Sample Size Calculations
Large differences in required sample sizes were observed for small effect sizes, with the largest differences being between methods in the AD group (Supplemental Table 7). However, with larger effect sizes (in line with expectations in clinical trials), differences in required sample size between the 2 methods became negligible for both SCD and AD (Supplemental Table 7).

Simulations
Simulations with 5% coefficient of variance showed results similar to those for the simulated time-activity curves obtained with almost no noise (0.05% coefficient of variance). Therefore, to mimic real cohort data, only the results from time-activity curves with a 5% coefficient of variance were reported.
Simulations revealed that under the SCD (almost no binding) and low-binding AD patient conditions, an inverse relation was observed; that is, with increasing flow, a decreasing bias for SUVr (with respect to true DVR) was observed (Fig. 4). A similar behavior was also observed under the medium-binding AD patient condition, but to a lesser extent. In the high-binding condition for AD patients, however, a relatively smaller effect of flow was observed on SUVr, implying that SUVrs remained relatively constant irrespective of the change in flow. In the case of DVR, no effect of flow was observed with any of the conditions (Fig. 4).
On the basis of simulations, percentage bias in SUVr with respect to the true DVR varied with the choice of SUVr scanning interval and the underlying binding condition (Fig. 5). In general, SUVr overestimated DVR for all simulated R 1 conditions from 80 min after injection; however, the impact of the change in flow on the directionality of the bias seems also to vary with respect to the choice of SUVr scanning interval (Fig. 5). DISCUSSION We compared changes in 18 F-flortaucipir specific binding using SUVr and DVR. In a 2-y longitudinal study, changes in 18 F-flortaucipir DVR and SUVr were comparable in all patient groups. Only small changes in R 1 occurred during this period, but these most likely contributed to the lack of difference between DVR and SUVr. However, simulations demonstrated marked differences between DVR and SUVr when large(r) changes in R 1 were introduced. In addition, these differences between DVR and SUVr were shown to be dependent on the underlying level of tau pathology.
The most important finding in this study was the lack of major differences in the percentage change between 18 F-flortaucipir DVR and SUVr in a 2-y observational study. Congruently, sample size calculations based on these data to inform future trials showed negligible differences between methods. Unlike a previous study using 11  , this finding indicates that 18 F-flortaucipir SUVr provides an accurate estimate of change in specific binding in both patient groups. There are several possible reasons for the differences in findings between the 2 studies. First, an important factor contributing to our findings could be the relatively small or nonexistent R 1 differences in this cohort. Previously, using 11 C-PiB (6), we reported larger R 1 changes in AD patients, which induced a large difference between SUVr and BP ND . This effect might perhaps indicate that 11 C-PiB is more sensitive to changes in R 1 than is 18 F-flortaucipir. However, flow sensitivity may also depend on the scanning interval relative to tracer kinetics, as was seen previously for 11 C-PIB (6). Similarly, this is the scenario for 18 F-flortaucipir, and we therefore cannot directly compare the 2 tracers in this respect. Second, it has been reported that accumulation of tau pathology is a slowly developing process, with annual percentage changes of about 0.5%-3% in Ab-positive cognitively unimpaired subjects and up to 3%-10% in Ab-positive cognitively impaired subjects (34)(35)(36)(37). The annual percentages change in the present study was generally comparable in SCD subjects (on average, 1.08% SUVr and 1.28% DVR) and slightly lower in AD  subject (2.73% SUVr and 2.52% DVR). The test-retest repeatability of 18 F-flortaucipir, as reported previously (38), lies at around 1.98% (0.78-3.58) for DVR and 3.05% (1. 28-5.52) for SUVr at 80-100 min. Although the test-retest repeatability was significantly better for DVR (38), annual percentage changes as found in the present study still fall within 1 SD of the test-retest repeatability for both DVR and SUVr, suggesting that observed changes might be too small to detect differences between analytic methods. Finally, differences with respect to tracer target affinity, isotope ( 11 C vs. 18 F), and pharmacokinetic behavior might have introduced differences that caused the differences in results.
Currently, the effects of pharmacotherapeutic interventions on cerebral blood flow are unclear. Therefore, we performed simulations to investigate the impact of large(r) changes in relative cerebral blood flow/R 1 on the accuracy of SUVr and DVR. The bias with SUVr relative to DVR was different for each flow condition, and this bias was additionally influenced by the underlying tau load, with decreasing bias in cases of low tau load/binding or constant bias for high tau load/binding. Depending on the underlying tau load, regional changes in flow resulted in variable changes in SUVr, which was not the scenario with DVR. Similar findings were previously observed using 18 F-cyclofoxy (39).
On top of flow condition and the underlying tau load, the choice of SUVr time interval also effected the accuracy, which was again different for different binding conditions. A previous study found large positive biases for SUVr using different time intervals when compared with dynamic methods (8). Furthermore, Golla et al. (8) observed that the bias in SUVr for a specific scanning interval is not constant but is dependent on the underlying tau load and the choice of SUVr scanning interval. This has important implications, since scanning intervals for static protocols are often not strictly enforced; thus, deviations in scanning intervals between static and longitudinal scans are common. These discrepancies will increase variability and uncertainty, which will increase required sample sizes for SUVr. Differing underlying tau load in the sample studied will only increase the bias in SUVr further. It is worth noting that, in the current study, SUVr was extracted from the dynamically acquired data. In addition, scanning interval was strictly enforced in the context of the 2 scanning sessions within the dynamic protocol. For both these reasons, SUVr in this study was not affected by deviations in scanning, and the results may therefore be too optimistic in this respect.
The discrepancies between methods using simulations may have important implications for longitudinal 18 F-flortaucipir studies and intervention studies. Our findings imply that SUVr is not the parameter of preference when large variations in blood flow are expected, although to what order of magnitude remains to be elucidated. A consideration to address when using repeated dynamic scans is potential selection bias, because severely affected patients might not be able to undergo such a demanding procedure. In patients with moderate to severe AD, this is indeed debatable. However, pharmacotherapeutic trials currently show a shift in target population, primarily including patients with mild, prodromal, or preclinical autosomal-dominant AD. Those patients can tolerate the longer dynamic scan procedures. 18 F-flortaucipir is useful for investigating pathologic tau load differences between SCD subjects and AD patients. However, in an early-dementia cohort for which we do not expect specific binding in the neocortex, measurement of tau deposition shows large variability. Indeed, in such a sample, 64% of the cortical signal variability can be explained by off-target binding (40). Partial-volume correction does not completely explain the variability in the cortical signal. Therefore, the variability in the signal in cohorts with low tau deposition related to off-target binding should be considered when examining early tau deposition using 18 F-flortaucipir.

CONCLUSION
Static scanning protocols provide accurate estimates of specific 18 F-flortaucipir binding in observational studies. Dynamic scanning protocols and fully quantitative data analysis methods are preferred when large(r) flow changes in the brain are expected (such as in later disease stages or pharmacotherapeutic interventions). Use of semiquantitative methods in such conditions carries the inherent risk that potential effective therapeutic interventions are discarded, especially when expected effect sizes are small.

DISCLOSURE
Research at the Amsterdam Alzheimer Center is part of the neurodegeneration program of Amsterdam Neuroscience; the Amsterdam Alzheimer Center is supported by Alzheimer Nederland and Stichting VUmc funds. 18   and Alzheimer Nederland. Bart van Berckel has performed contract research for Rodin, IONIS, AVID, Eli Lilly, UCB, DIAN-TUI, and Janssen; was a speaker at a symposium organized by Springer Healthcare; has a consultancy agreement with IXICO for the reading of PET scans; is a trainer for GE; and receives financial compensation only from Amsterdam UMC. No other potential conflict of interest relevant to this article was reported.

ACKNOWLEDGMENTS
We kindly thank all participants for their contribution. We thank Ronald Boellaard for sharing his knowledge and thoughts about the project.

KEY POINTS
QUESTION: How do the semiquantitative (SUVr) and quantitative (R 1 , BP ND ) parameters of longitudinal 18 F-flortaucipir PET scans, and their vulnerability to changes in blood flow, compare in subjects along the AD continuum?
PERTINENT FINDINGS: In a 2-y longitudinal 18 F-flortaucipir PET study including 38 subjects with SCD and 24 patients with AD, relative cerebral blood flow changes (R 1 ) were small, and semiquantitative (SUVr) and quantitative (BP ND ) parameters yielded highly similar estimates of specific binding. However, simulations showed that large(r) flow changes may potentially affect 18 F-flortaucipir SUVr.

IMPLICATIONS FOR PATIENT CARE:
Given that it is currently unknown to what order of magnitude pharmacotherapeutic interventions may induce changes in cerebral blood flow, caution may be warranted when changes in flow are large(r), and DVR/BP ND may be preferred under such conditions to ensure representative quantification of 18 F-flortaucipir PET images.