Participants
The study included N = 121 participants from the AMYloid Imaging to Prevent Alzheimer’s Disease (AMYPAD) Prognostic and Natural History Study (PNHS) [15], who were scanned across four different centres: The University of Edinburgh (UEDIN), Barcelona βeta Brain Research Center (BBRC), Amsterdam University Medical Centre, location VUmc (Amsterdam UMC, VUmc), and University of Geneva (UNIGE). All participants had at least one PET and one T1-weighted MR scan available, and visual assessment of the PET scans was performed locally by a trained nuclear medicine physician according to the manufacturer’s reading guidelines [1, 2]. Readers were blinded to the quantitative outcome measure of the scans. In addition, participants underwent standard neurological screening and neuropsychological assessment (e.g. mini-mental state examination, MMSE) and information regarding their APOE-ε4 status was available. Before participating in the study, all participants provided written informed consent in accordance with the Declaration of Helsinki. Study protocols were approved by all local Medical Ethics Review Committees. EudraCT Number: 2018-002277-22, registered on: 25-06-2018, https://www.clinicaltrialsregister.eu/ctr-search/trial/2018-002277-22/NL.
Image acquisition
Prior to each PET scan, an MR sequence (MRAC) or low-dose CT (ldCT) was acquired for attenuation correction purposes (Additional file 1: Table S1 specifies which centres used an MR and which a CT scanner). Participants then received a bolus injection of either [18F]flutemetamol (N = 90, 186 ± 10 MBq) or [18F]florbetaben (N = 31, 283 ± 20 MBq) and underwent a dynamic PET scan according to a dual-time window protocol [16]. This scanning protocol consisted of an early dynamic scan from 0 to 30 min post-injection (p.i.) followed by a break of 60 min, and then a late dynamic scan from 90 to 110 min p.i. For each scanner, recommended clinical reconstruction settings for that scanner were used and scans were reconstructed into 18 frames (6 × 5, 3 × 10, 4 × 60, 2 × 150, 2 × 300, and 1 × 600 s) for the early scan, and four frames (4 × 300 s) for the late scan. All sites applied the following corrections: detector normalization, dead time, attenuation, scatter, decay, and randoms. Site-specific reconstruction methods can be found in Additional file 1: Table S1.
Image QC and processing
First, a quality control (QC) check of the image meta-data was performed (i.e. frame start times, duration, and the number of frames) and the presence of motion between PET scans, and corresponding MRAC or ldCT, was evaluated. In the case of severe motion (> 5 mm), these scans were excluded from analyses. In addition, it was verified whether the whole brain was in the field of view and whether no technical errors occurred, such as a delay in acquisition start time. Next, presence of between-frame motion was assessed visually. Then, both early and late phases of the PET scan were coregistered to the T1-weighted MR scan using rigid registration and the resulting PET scans were resampled to have the same voxel size as the T1-weighted MR. This was done to prevent changes to the predefined region of interest (ROI) template images, which were already in the same space as the T1-weighted MR scan. Following this step, coregistered PET images were combined and the resulting scan was divided into five blocks of 10 min duration (A: frames 1–15, B: frames 16–17, C: frame 18, D: frames 19–20, and E: frames 21–22). Presence of motion between blocks was assessed visually, by comparing whether contours between blocks were overlapping. Next, using rigid co-registration (based on Elastix software), blocks were co-registered and the registration parameters were saved [17, 18]. Based on a visual assessment and the registration parameters, it was determined whether motion exceeded the maximum allowed threshold (1 mm translation in each direction) and should be corrected for (using Elastix-based rigid co-registration). Subsequently, subject-specific regions of interest generated using Learning Embeddings for Atlas Propagation (LEAP) [19] (in T1-weighted MR space) were applied to the combined PET scan to extract a reference tissue time-activity curve (TAC) of the cerebellar grey matter. As regions of sufficient volume were used (> 5 ml), no partial volume corrections were performed.
Parametric analysis
Interpolation of the missing data points of the reference tissue TAC (corresponding to the 60 min break) was performed using the reversible two tissue compartment model (4-rate constants) with additional blood volume fraction parameter (2T4k_Vb) and a tracer-specific plasma input function as described previously [16]. Visual QC of the interpolated reference tissue TAC was performed to assess whether the interpolation of the missing data points had a smooth connection with the measured data. In the case of sub-optimal interpolation (i.e. clear discontinuity between interpolated and measured data), cubic interpolation was used as alternative (N = 8). Next, parametric modelling was performed using the basisfunction-based implementation of the two-step simplified reference tissue model (SRTM2) [20], as implemented in the PPET software package [21] and validated previously [22], to compute parametric BPND and relative tracer delivery (R1) images. For SRTM2, k2’ was determined by taking the median k2’ across all voxels with a BPND higher than 0.05 from a first run using receptor parametric mapping (RPM) [23, 24]. In addition, SUVR was calculated from 90 to 110 min p.i. To allow for comparability between metrics, DVR was calculated as BPND + 1. Finally, the subject-specific global cortical average (GCA) template, which is a composite region consisting of frontal, temporal, parietal, and insular cortices, precuneus, and striatum [25], was applied to the parametric data to obtain global SUVR and DVR values. In addition, the subject-specific ROIs generated using LEAP were used to extract values for regions that are known to show early accumulation of Aβ plaques: precuneus, posterior cingulate cortex (PCC), and orbitofrontal cortex (OFC) [12, 26, 27].
Statistical analysis
All statistical analyses were performed in R (version:4.0.3; R Foundation for Statistical Computing, Vienna, Austria) and stratified per tracer. A statistically significant result was defined as p < 0.05; no corrections for multiple comparisons were applied.
First, demographic differences between Aβ-positive and negative participants (based on a visual assessment) were investigated using t tests (or in the case of non-normal distribution, Mann–Whitney U-tests) and Chi-square tests.
Relationship between SUVR and DVR
Linear regression analyses were used to assess the relationships between SUVR and DVR. In addition, Bland–Altman analyses [28] were used to assess potential bias between these metrics, and the presence of proportional bias was determined visually for the GCA and three early regions. When present, proportional bias was further evaluated by fitting a regression line through the Bland–Altman plot.
Relationship between (bias in) SUVR and relative CBF
First, bias in SUVR relative to DVR (SUVRbias) was calculated ((SUVR-DVR)/DVR*100%). Then, generalized linear models (GLMs) were constructed to understand whether the relative tracer delivery (R1) could explain the remaining variance of either SUVR or SUVRbias beyond the main predictor, DVR. Note, the covariate “centre” is only included for [18F]flutemetamol.
$$\begin{aligned} & {\text{SUVR}} \sim {\text{DVR}} + R_{1} + {\text{DVR}}*R_{1} + {\text{Age}} + {\text{Sex}} + {\text{APOE}} - \varepsilon 4 + {\text{centre}} \\ & {\text{SUVR}}_{{{\text{bias}}}} \sim {\text{DVR}} + R_{1} + {\text{DVR}}*R_{1} + {\text{Age}} + {\text{Sex}} + {\text{APOE}} - \varepsilon 4 + {\text{centre}} \\ \end{aligned}$$
Before applying the GLMs to the data, all predictors (DVR, R1, and DVR*R1) and covariates (age, sex, APOE-ε4 carriership, and centre in the case of [18F]flutemetamol) were correlated with each other to check for collinearity. This was done using Pearson’s correlation, point-biserial correlation, and Goodman Kruskal’ lambda for continuous, continuous and categorical or categorical variables, respectively. In the case of high overlap (r ≥ 0.70) [29], the variables were correlated with the dependent variable (SUVR and SUVRbias) to determine which one should be deleted from the model to remove redundancy, i.e. the one with the lowest correlation. All analyses with the final model were performed using the GCA, as well as with the three early regions.