Background
Positron emission tomography (PET) imaging has played an integral role in Alzheimer’s disease (AD) neuroimaging research by yielding precise in vivo measurement of β-amyloid (Aβ) pathology [
1]. Still, Aβ PET studies are limited by the variability that can be introduced through non-specific binding properties of radiotracers [
2], scanner and reconstruction differences, and variations in analysis pipelines [
3,
4]. This variability ultimately frustrates efforts to combine data for meta-analyses and multicenter studies [
5], track longitudinal changes in Aβ burden [
6], and establish universal cut points for Aβ positivity [
5]. In light of these issues, Klunk et al. [
5] developed the Centiloid (CL) method, which standardizes total Aβ burden assessed with PET imaging agents by (1) establishing a standard analysis pipeline for quantifying cortical standardized uptake value ratios (SUVRs) and (2) converting SUVRs across various Aβ radiotracers and analysis methods to a common scale. In other words, investigators can choose to express their data in CL units either by using the standard CL analysis pipeline or by cross-calibrating their data against previously validated data, ultimately yielding linear equations for transforming their data into CL units.
The Alzheimer’s Disease Neuroimaging Initiative (ADNI) is an ongoing, multisite observational study of AD. Presently, thousands of [
18F]florbetaben (FBB; 90–110 min) and [
18F]florbetapir (FBP; 50–60 min) SUVRs are available to the scientific community through ADNI’s database website (
http://adni.loni.usc.edu), with more anticipated to be collected and shared under the current protocol (ADNI-3). Furthermore, ADNI-compatible PET acquisition and processing methods are being implemented in other ongoing multisite AD studies such as the Longitudinal Early-Onset Alzheimer's Disease Study (LEADS) [
7] and the Standardized Centralized Alzheimer’s and Related Dementias Neuroimaging (SCAN) project [
8]. Data from ADNI and other studies with compatible protocols are also frequently used by the scientific community. As of 2020, ADNI data have been downloaded more than 100 million times by users across workforce sectors [
9], with ADNI additionally having been credited in over 1800 scholarly publications. Taken together, and in conjunction with the Aβ PET issues described above, a standard outcome measurement like the CL method holds the potential to benefit a large number of PET studies; this would enable harmonization of Aβ PET across various sites, scanners, and tracers, ultimately increasing statistical power for research studies and clinical trials that wish to use Aβ PET imaging for prediction or outcome measures.
To promote such harmonization, the primary objectives of this work were to (1) create direct CL conversion equations for FBB and FBP SUVRs derived from ADNI processing methods to facilitate both cross-sectional and longitudinal analyses and (2) establish positivity thresholds for ADNI-derived FBB and FBP SUVRs based on independent young control samples. We accomplished our first objective by analyzing FBB and FBP datasets available on the Global Alzheimer’s Association Interactive Network (GAAIN) website (
www.gain.org/centiloid-project) through ADNI’s automatic, native-space pipeline and subsequently following procedures necessary for level-2 analysis as described by Klunk et al. [
5]. While linear equations for converting FBB and FBP SUVRs derived from the standard CL pipeline to CL units have been published [
10,
11], these transformations are only valid when applied to data analyzed with the standard CL analysis pipeline; PET processing pipelines that use different analysis approaches, such as that of ADNI, change the quantification parameters and therefore the relationship of analysis outcomes to CL units. We defined CL conversions for ADNI-derived FBB and FBP SUVRs normalized to the whole cerebellum, for use in cross-sectional analyses, and normalized to a composite reference region (made up of eroded subcortical white matter, brainstem, and whole cerebellum), which has shown greater reliability for longitudinal analyses [
12‐
15].
To accomplish our other primary objective, we used the CL conversion equations to calculate thresholds in CL units for FBB and FBP that are compatible with ADNI acquisition processing methods. For FBB, existing thresholds have been reported based on the detection of Aβ [
16] or the separation of Aβ-positive patients and Aβ-negative controls [
17,
18] but these thresholds were defined using pipelines that differ from that of ADNI, so they are not directly applicable to our data. In addition, these thresholds may be less sensitive to early increases in Aβ burden than techniques that emphasize the detection of Aβ relative to young individuals with no evidence of any Aβ. Thus, in order to define a FBB threshold that is congruent with the use of a young control sample as a standard of comparison, we examined FBB uptake in young healthy controls and validated this finding with a data-driven approach in ADNI participants using the updated ADNI PET pipeline. For FBP, we used a threshold (1.11) that was based on the upper 95% confidence interval of mean cortical FBP SUVRs relative to the whole cerebellum in a young control group [
19] and transformed to the ADNI FreeSurfer (FS) pipeline initially using FS v5.3 [
3]. Here, we validate this FBP threshold with the updated FS v7.1-based pipeline.
Discussion
Variability across radiotracers, scanners, and analysis pipelines limits the comparison of Aβ PET measurements across studies. To address these issues and facilitate comparison of Aβ burden across different ligands, we used a variety of datasets to derive CL transformations for ADNI FS v7.1 FBB and FBP SUVRs. We applied these CL transformations to validate FBB and FBP thresholds that were derived using congruent approaches and that can be used to determine Aβ positivity in datasets that include both tracers. While the primary objectives of this work were to calculate the CL conversions and thresholds for ADNI FS v7.1-derived SUVRs normalized to the whole cerebellum (as is done for cross-sectional studies), we additionally defined CL conversions for SUVRs normalized to a composite reference region, which is recommended for longitudinal studies using FBP. All of these findings are immediately applicable to the thousands of current and anticipated SUVRs acquired using methods compatible with ADNI. More broadly, our CL transformations can be applied to any FBB and FBP PET image data that were (1) collected according to ADNI-like acquisition and pre-processing protocols and (2) analyzed using the updated PET pipeline described here, which is based on a native-space, FS-based quantification approach.
The FBP and FBB thresholds were based on the upper limit of cortical uptake in whole cerebellum-normalized SUVRs in young control samples. For FBP, we confirmed that our previously validated threshold (1.11, 20 CL) that was transformed from Avid-acquired young control data [
3,
19] was unchanged using the updated ADNI pipeline. For FBB, we analyzed young control data acquired by LMI [
20] with the updated ADNI pipeline to derive a threshold (1.08, 18 CL for SUVRs normalized to the whole cerebellum) and verified a similar result using a data-driven approach with the existing ADNI FBB sample (1.10, 21 CL for SUVRs normalized to the whole cerebellum). While these values are similar, the 1.08 (18 CL) threshold is preferred as it was derived in a sample independent from ADNI and because it is based on healthy individuals who are free of Aβ burden, making it methodologically congruent with the FBP threshold. Previous studies have developed FBB thresholds using histopathological confirmation of Aβ in postmortem brain tissue and other methods that have primarily focused on the detection of clinical characteristics and/or Aβ [
16,
17,
23]. These thresholds may be less sensitive to early elevations in Aβ burden compared to the approach used here, which relies on cortical Aβ and associated Aβ variability in healthy controls with no Aβ burden. In addition, many previous FBB studies have used the cerebellar cortex as a reference region, resulting in lower reference region estimates and a higher cortical Aβ SUVR threshold, as opposed to the whole cerebellum (white and gray matter), which results in higher reference region estimates and a lower cortical SUVR as reported here.
Previous studies have demonstrated that both FBB and FBP SUVRs are appropriate for conversion to CL units using the standard CL analysis pipeline [
5,
10,
11]. The analysis presented in this work confirms these findings. In both FBB and FBP cohorts, we observed a strong correlation between ADNI FS v7.1 pipeline-derived whole cerebellum-normalized SUVRs and their respective standard CL pipeline-derived PiB SUVRs (
R2=0.958 for FBB;
R2=0.897 for FBP).
The equation describing the linear regression of ADNI FS v7.1 FBB SUVRs normalized to the whole cerebellum and the standard CL pipeline PiB SUVRs (Eq.
3) indicates that FBB has a narrower dynamic range compared to PiB (slope<1), which is consistent with the previous literature [
24]. The increased variance we observed in the FBB CL units relative to the PiB CL units is likely to be due to differences between tracers. Rowe et al. [
10] similarly found increased variance in FBB CL units compared to PiB CL units when SUVRs for both tracers were derived from the standard CL pipeline. However, compared to our findings, this group reported less precision in FBB CL units (SD=6.81) and a greater standard deviation ratio (1.96), suggesting that compared to the CL pipeline, the ADNI pipeline introduces less variance.
FBP has previously been reported to have about one-half of the dynamic range of PiB [
3]. The slope of the equation describing the linear regression of ADNI FS v7.1 FBP SUVRs normalized to the whole cerebellum and CL pipeline PiB SUVRs (Eq.
4) reflects this. We also found more variability in FBP CL values in young controls compared to PiB CL values. Navitsky et al. [
11] also found greater variance in CL units converted from standard CL pipeline-derived FBP SUVRs compared to those from PiB. This study also reported larger SD (12.07) and standard deviation ratio (3.96) for YC CL units when compared to the present work, so the increased variance that we observed is likely due to differences between FBP and PiB as the ADNI pipeline appears to introduce less variance than the standard CL pipeline. These findings together reinforce the fact that conversion to CL values does not improve the precision of [
18F] tracers relative to PiB.
The equations that describe whole cerebellum-normalized SUVR-to-CL units for both FBB and FBP (Eqs.
5 and
6) are markedly different from those previously published for the same tracers [
10,
11]. Importantly, such work has transformed SUVRs derived from the standard CL pipeline to CL units whereas we transformed SUVRs derived from FS v7.1. Sizes of the cortical regions between the two pipelines differ (Supplemental Table
2), as does sampled tissue (Supplemental Figure
2). The equations describing conversion from composite region-normalized SUVR-to-CL units (Eqs.
7 and
8) are even more different from previously published transformations, likely due to both the different cortical and reference regions. While the use of a reference region containing subcortical white matter reduces the dynamic range of amyloid PET SUVRs, several studies have demonstrated that it optimizes longitudinal PiB and FBP reliability [
12‐
15]. It is important to note that the transformation equations described in this manuscript are not appropriate for converting any other type of SUVR to CL units; they only are only suitable for FBB and FBP outcomes derived from the ADNI FS v7.1 pipeline using either the whole cerebellum or composite as reference region.
It is worth noting that in the GAAIN datasets, relative standard deviations of FBP-to-PiB CL units are more than two times greater than that of FBB-to-PiB CL units. The age distribution is comparable between the two cohorts [
10,
11], and the mean and SD of PiB SUVR and CL units are almost identical between young controls. However, the standard deviation ratio of whole cerebellum-normalized SUVRs in the FBP cohort is 67% greater than that of the FBB cohort. Thus, the disparity in standard deviation ratios between FBB and FBP CL units may be attributed to differences in dynamic ranges relative to PiB. Alternatively, the increased standard deviation ratio of FBP may be due to scanner differences, as FBP CL data was collected on three different PET scanners at three different sites [
11], whereas FBB CL data was collected on two PET scanners at a single site [
10].
Our thresholds expressed in CL units are consistent with recent reports for FBB [
18,
25] and FBP [
26] specifically, and for Aβ burden studies in general [
25‐
27]. Such studies have described CL thresholds that range from 12 to 35 depending on the stringency of the threshold standard, quantification approach, and tracer; lower thresholds emphasize early detection and higher thresholds maximize specificity [
18]. Using positive visual reads as the standard, our CL thresholds (FBB 18 CL; FBP 20 CL) fall slightly lower than the published range of 25–35 [
25]. However, using CSF [
26] and the presence of plaques with histopathological examination as the standard, our thresholds are on the high end of the reported ranges. That is, in one neuropathological validation study, Rowe et al. [
25] reported a range of 15–20 CL with FBB or PiB. Additionally, La Joie et al. [
27] reported a threshold of 12 CL using PiB for detection of moderate to frequent plaques and 24 CL for identification of AD neuropathologic change (a composite score). It should also be noted that in another study that used an MRI-based quantification approach similar to ours, Dore et al. [
18] reported a range of 22–28 FBB-derived CL and thus, our range is comparatively lower.
A strength of the present study is its multi-scanner nature, since ADNI pre-processing allows data from multiple sites to be merged despite differences in scanners, reconstruction methods, and spatial resolution.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.