Estimation of amyloid distribution by [¹⁸F]flutemetamol PET predicts the neuropathological phase of amyloid β-protein deposition

The subject cohort of 97 cases was included in the [¹⁸F]flutemetamol efficacy analysis of the GE067-007 (ClinicalTrials.gov identifier: NCT01165554) and GE-067-026 (ClinicalTrials.gov identifier: NCT02090855) phase 3 end-of-life clinical trial and autopsied after death (Suppl. Tab. 1) [6, 17]. The criteria for the selection of these 97 cases from all 106 studied subjects included in the trial were the availability of necessary parameters for this study, i.e., Aβ-phases, comparable measurements for cortical and caudate nucleus tracer retention. For 9 cases caudate SUVRs could not be obtained for the reasons listed in Suppl. Tab. 2. Dementia, defined according to the DSM IV criteria, was noted as present or absent. This was a phase 3 multicenter PET study of [¹⁸F]flutemetamol injection for detecting brain Aβ. Local institutional review boards or ethics committees approved the study protocol before initiation. All subjects or their legal representatives gave prior written informed consent/assent. Consecutive eligible subjects were ≥ 55 years of age, terminally ill with a life expectancy < 3 years, with general health adequate to undergo study procedures. Patients died of natural causes and no serious adverse events were attributable to [¹⁸F]flutemetamol injection [11]. Subjects were ineligible if they were pregnant/lactating, had known/suspected structural brain abnormalities, contraindication(s) for PET, known/suspected hypersensitivity/allergy to [¹⁸F]flutemetamol injection (or any component), or had participated in any clinical study using an investigational product within 30 days of signing consent.

The scan-to-death time intervals ranged between 0 and 846 days (mean 215 days; median 154 days).

Brain material received at autopsy and previously used for diagnostic purposes supporting the GE067-007/GE-067-026 phase 3 clinical trials was examined. All brains were formalin-fixed. The brains were cut in coronal slices and screened macroscopically. For histopathological analysis and for assessing the amounts of AD-related amyloid plaques, NFTs, and neuritic plaques, we examined paraffin-embedded tissue including parts of the frontal, parietal, temporal, occipital and entorhinal cortex, the hippocampal formation at the level of the lateral geniculate body, basal ganglia, thalamus, amygdala, midbrain, pons, medulla oblongata, and cerebellum. Paraffin sections of 5 µm thickness from all blocks were stained with hematoxylin & eosin and anti-Aβ antibodies (anti-Aβ; 4G8, SIG-39220, Covance, USA, 1:100, formic acid and heat pretreatment). For neuropathological diagnosis sections were stained with the Bielschowsky silver method and immunohistochemical methods for abnormal phosphorylated τ-protein (anti-human PHF-Tau monoclonal antibody; AT8, Prod. No. MN1020, 1:40, Thermo Scientific, United Kingdom), α-synuclein (anti-α-synuclein monoclonal antibody, Prod. No. NCL-L-ASYN, Lot No. 6005209, 1:40, Leica Microsystems, United Kingdom), and ubiquitin (anti-ubiquitin polyclonal antibody, Prod No. Z0458, 1:400, DakoCytomation, United Kingdom). Primary antibodies were detected with biotinylated secondary antibodies (DakoCytomation E0354, UK) and visualized with the DABMap Kit (Ventana, USA). The phase of Aβ-plaque pathology (Aβ-phase) was assessed after screening the Aβ-stained sections for plaque distribution according to previously published protocols [3, 36]. The neuropathological diagnosis of AD-pathology was performed as recommended [16].

Demented cases with at least intermediate NIA-AA degrees of AD-pathology [16] were considered as symptomatic AD cases; non-demented individuals with AD-pathology were referred to as p-preAD cases [28, 34, 38]. Non-demented cases without AD-pathology were classified as non-AD controls. Non-AD controls included cases with NFT-pathology in the medial temporal lobe termed primary age-related tauopathy (PART) [10]. Patients with non-AD dementia encompassed demented patients with vascular dementia, Lewy body disease (LBD) and frontotemporal lobar degeneration with τ-pathology (FTLD-tau: argyrophilic grain disease, NFT-predominant dementia, Pick’s disease) that did not exhibit intermediate or high degrees of AD-pathology indicating that AD-pathology was presumably not responsible for dementia. NFT-predominant dementia was classified according to the recommended classification for FTLD [25] as a distinct form of FTLD-tau and was, therefore, not considered as PART although there is an overlap between PART cases with dementia and NFT-predominant dementia [10, 19].

Amyloid PET imaging was performed at 12 different imaging sites in the USA and in Europe [31, 34]. Before PET imaging, subjects underwent head CT or magnetic resonance imaging (MRI), unless prior images (obtained within 12 months) were available. [¹⁸F]Flutemetamol injection was administered intravenously at a dose of 185 or 370 MBq of radioactivity at physician discretion [34]. PET images were acquired in 2-min frames on PET/CT cameras, beginning approximately 90-min post-injection, which was attenuation corrected using CT data. Frame-to-frame motion correction was performed on the dynamic data before the frames were averaged to give a 10–20 min scan. Equipment used to capture images varied across the 12 imaging sites [34]. Most images were reconstructed iteratively to form 128 × 128 axial slices, and a Gaussian post-reconstruction smoothing filter was applied to some to achieve uniform image resolution across sites.

The [¹⁸F]flutemetamol uptake was measured for six gray matter volumes of interest (VOIs) and adjusted for atrophy manually, covering the anterior cingulate, the prefrontal cortex, the lateral temporal cortex, the parietal cortex, one VOI covering both posterior cingulate and precuneus, and one subcortical VOI in the head of the caudate nucleus [6, 34]. Quantitative SUVR calculations were made using pons as reference region [22]. A global neo- and allocortical, composite SUVR (SUVRcort, Table 1) was calculated from SUVRs obtained from anterior cingulate, prefrontal, lateral temporal, parietal, and posterior cingulate cortex including the precuneus region [39]. The SUVR for the caudate nucleus (SUVRcaud, Table 1) was determined based on VOI measurements of both the left and right caudate nucleus (anterior aspect). The caudate VOIs were drawn on a para-sagittal plane which intersected the thalamus, internal capsule, caudate head and frontal white matter. Image processing and VOI analysis was performed using VOIager 4.0.7 (GE Healthcare, Uppsala, Sweden).

To allow staging of amyloid PET Aβ-pathology progression as suggested by Hanseeuw et al. [15], Grothe et al. [13] and Cho et al. [8] (Table 1), visual interpretation of the PET/CTs was carried out by 5 independent raters, who were experienced in PET analysis. The overall presence/absence of amyloid as well as the presence/absence of striatal amyloid was assessed (Suppl. Tab. 1). Striatal amyloid was, thereby, visually assessed in the head of the caudate nucleus and the putamen [6]. For this procedure with [¹⁸F]flutemetamol an inter-rater read agreement of at least 80% has been previously published [6, 11, 17]. For determination of the PET-Aβ stages [15] we used the majority reads.

Linear regression analysis was calculated using SPSS 24 statistical software (IBM, Armonk, NY, USA). For this purpose the Aβ-phase, which was determined by neuropathological analysis, was set as the dependent variable and SUVRcort, SUVRcaud, SUVRcort + caud, as well as the stages of anatomical distribution of amyloid PET-tracer retention, called the PET-Aβ stages, and the new parameter identified in this study describing threshold levels that determine estimates for the underlying pathological Aβ-phase, i.e., the PET-Aβ phase estimate (see Table 1), were set as independent variables in linear regression model terms. In a first analysis the linear regression model only included the dependent and one independent variable. In the event that associations were found we used a second model term in which the independent variable was controlled for age, sex, and last scan-to-death interval. In addition to R² and p values the standardized β coefficient was determined.

To clarify whether it is better to use the cortical SUVRs obtained in five different regions separately or just their composite, i.e., SUVRcort, we first analyzed by ANOVA using the Games-Howell post-hoc test to correct for multiple testing whether differences exist among the five cortical SUVRs and the composite SUVRcort in distinguishing between the Aβ-phases. Since there were no major differences in distinguishing between the Aβ-phases among these SUVRs (Suppl. Tab. 3) we decided to continue with SUVRcort to represent cortical Aβ-deposition.

To determine thresholds for distinguishing the different neuropathologically defined Aβ-phases by SUVR-levels we determined the mean, median, variance, and range of SUVRcort and SUVRcaud for each Aβ-phase including Aβ-phase 0. To identify the cutting point between brains without detectable Aβ-plaques and with Aβ-pathology we identified the maximum SUVRcort in Aβ-phase 0 and the maximum SUVRcaud in Aβ-phases 0–2 (i.e. in the absence of plaques in the striatum). These maximum SUVRs for amyloid plaque free cortex and caudate nucleus were rounded to the next decimal place to indicate that all SUVRs equal or higher indicate the presence of amyloid. To distinguish between amyloid phases 1–5 we first carried out an ANOVA analysis corrected for multiple testing with a Games-Howell post-hoc test. In the event that a significant increase in SUVR was found when comparing cases with Aβ-phases 0 and 1, 1 and 2, 2 and 3, 3 and 4, or 4 and 5, we decided to identify thresholds for distinguishing the respective Aβ-phase. For determining these thresholds we calculated the mean of the two respective mean SUVRs, e.g., that of the means from Aβ-phases 1 and 2, 2 and 3, 3 and 4, or 4 and 5, respectively, and rounded it to one decimal place by considering the variances: in the event that the cases of the higher Aβ-phase showed a higher variance in SUVRs than those of the lower Aβ-phase then the threshold was determined by rounding to the lower decimal place whereas in the opposite constellation the threshold was determined by rounding the mean of the mean SUVRs to the higher decimal place.

After applying the thresholds and predicting Aβ-phases by PET the percentage of correctly classified cases was calculated as well as that of cases being correctly classified or classified one phase higher or lower than seen at pathological observation. Since we were not able to distinguish between Aβ-phases 0–2 by PET these three phases were considered as one group for calculating the reclassification. Correlation between the neuropathological Aβ-phase and the PET-Aβ phase was calculated with 95% confidence interval (CI). A bootstrap method was performed to resample the data 1000 times using equal probability unrestricted random sampling with replacement and a 95% CI was calculated as a sensitivity analysis to validate the robustness of the estimates.

Linear regression analysis of the relationship between the Aβ-phases as the dependent variable and SUVRcort as independent variable revealed an association with R² = 0.584 (p < 0.001, β = 0.764) (Fig. 1a). A model term controlled for age, sex and scan-to-death time interval confirmed this association (p < 0.001, β = 0.719). A higher R² was observed for the correlation between the Aβ-phase as dependent variable and SUVRcaud as independent variable (R² = 0.670; p < 0.001, β = 0.818) (Fig. 1b). This association was also confirmed when controlling the independent variable for age, sex and scan-to-death interval (p < 0.001, β = 0.771). Combining both cortical and caudate SUVRs by simple addition as independent variable did not increase R² indicative for the association between the two parameters (R² = 0.666 instead of R² = 0.670 for SUVRcaud alone; p < 0.001, β = 0.816) (Fig. 1c). The same held true for the standardized β-coefficient in linear regression analysis controlled for age, sex and scan-to-death interval (p < 0.001, β = 0.772).

The currently published amyloid PET-based staging approaches for Aβ-pathology rely on the measurement/rating of cortical and subcortical (basal ganglia)-related amyloid positivity [8, 13, 15] and the hierarchical involvement of these regions as also known from histopathological studies [36]. Here, we used the majority binary reads (positive/negative) for overall amyloid pathology to define whether there is amyloid in the brain or not. Negative images (those without visible [¹⁸F]flutemetamol retention) were considered to represent PET amyloid stage 0. Positive images showed at least PET amyloid stage 1 (cortical amyloid deposition). PET amyloid stage 2 was considered when the striatal majority reads indicated amyloid positivity in the striatum as well (Table 1, Suppl. Tab. 1) according to the staging system published by Hanseeuw et al. [15] being compatible with those published by Grothe et al. [13] or Cho et al. [8], which are also based on the subsequent deposition of Aβ first in the cortex and later in the basal ganglia.

Regression analysis between the Aβ-phases as dependent variable and the PET amyloid stages revealed a strong association (R² = 0.697, p < 0.001, β = 0.835) (Fig. 1d), which was confirmed by linear regression analysis controlled for age, sex and scan-to-death interval (p < 0.001, β = 0.797).

To identify thresholds between the Aβ-phases we first calculated the range, mean, median, and variance of SUVRcort and SUVRcaud for cases within each Aβ-phase, respectively (Table 2). The range of SUVRcort in cases without Aβ-plaques (= Aβ-phase 0) has been considered as the normal range. The maximum SUVRcort in Aβ-phase 0 cases was 0.489 when using pons as reference region. By rounding to the next decimal place we considered cases with SUVRcort ≥ 0.5 as cases with detectable Aβ-pathology, i.e., Aβ-phase 1 or higher. Likewise, the range of SUVRcaud in cases with Aβ-phases 0–2, i.e., in phases when no plaques were detectable in the striatum, showed a maximum SUVRcaud of 0.596 (using pons as reference region). By rounding to the next decimal place, the threshold for SUVRcaud ≥ 0.6 was set to distinguish cases with and without caudate nucleus Aβ-pathology in addition to the detection of cortical Aβ by SUVRcort. To distinguish the different higher Aβ-phases we compared the mean and median SUVRcorts and SUVRcauds among the Aβ-phases and determined significant differences in the SUVRs of two subsequent Aβ-phases by ANOVA (Table 2). The means of Aβ-phases 0–2 did not vary significantly and were in the range of Aβ-phase 0 cases. Therefore, the first phase that was detected with [18F]flutemetamol amyloid PET was Aβ-phase 3. After phase 3 there was a continuous increase in the mean SUVRcort or SUVRcaud between Aβ-phase 3 and 4 whereas the difference between Aβ-phases 4 and 5 was more pronounced in SUVRcaud than in SUVRcort. The threshold distinguishing Aβ-phase 3 and 4 was oriented on the mean between the means of phase 3 and phase 4 cases. For SUVRcort this mean was 0.64 and for SUVRcaud 0.77. Since the variance of the Aβ-phase 3 SUVRs was lower than that of the Aβ-phase 4 SUVRs we rounded the thresholds to SUVRcort ≥ 0.6 for Aβ-phase 4 and higher and to SUVRcaud ≥ 0.7. The SUVRcort between Aβ-phases 4 and 5 showed major overlap in the ranges (Table 2). Therefore, we did not consider this parameter for the distinction between Aβ-phases 4 and 5. SUVRcaud, on the other hand, showed an obvious difference (Table 2) between Aβ-phases 4 and 5. The mean of the Aβ-phase 4 and 5 means was 0.96. Because the variance of Aβ-phase 5 SUVRcauds was lower than that of Aβ-phase 4 cases the cutting line for the phase 4/5 threshold was rounded up to SUVRcaud ≥ 1. Taken together, the threshold to distinguish Aβ-phases 3 from Aβ-phases 0–2 was either SUVRcort ≥ 0.5 and/or SUVRcaud ≥ 0.6. In comparison to Aβ-phase 3 cases, Aβ-phase 4 cases were best characterized by SUVRcort ≥ 0.6 and/or SUVRcaud ≥ 0.7 and the threshold for Aβ-phase 5 cases was identified as SUVRcaud ≥ 1.0, whereas SUVRcort did not contribute to the distinction between Aβ-phases 4 and 5 (Figs. 1a–e, 2; Table 2). Using these thresholds, four discrete PET-based estimates of the neuropathological Aβ-phase were created, termed PET-Aβ phase estimates (Fig. 2), with PET-Aβ phase estimate 0 = Aβ-phases 0–2, PET-Aβ phase estimate 1 = Aβ-phase 3, PET-Aβ phase estimate 2 = Aβ-phase 4, and PET-Aβ phase estimate 3 = Aβ-phase 5. However, one Aβ-phase 1 and two Aβ-phase 2 cases showed PET-Aβ phase estimate 1 indicating that few Aβ-phase 1 and 2 cases could already be diagnosed as positive for initial plaque pathology with amyloid PET by applying SUVR thresholds, but could not be separated from Aβ-phase 3 cases.

Accordingly, we categorized all cases of our sample using SUVRcort and SUVRcaud into PET-Aβ phase estimates. The regression analysis showed a strong association between the PET-Aβ phase estimate as the independent variable and the Aβ-phase as dependent variable (R² = 0.759, p < 0.001, β = 0.871) (Fig. 1e) providing the best relationship between a PET-derived parameter and the real neuropathological distribution of amyloid plaques as described by the Aβ-phases. This finding was in line with that of linear regression analysis controlled for age, sex and scan-to-death interval (p < 0.001, β = 0.840). The robustness of the estimates in the original study sample was confirmed through the bootstrap analysis. Correlation between PET-Aβ phase and Aβ-phase in the original sample was r = 0.88 (95% CI: 0.82, 0.92) and in the bootstrap analysis it was r = 0.88 (95% CI: 0.81, 0.92).

Using the PET-Aβ phase estimate to predict the Aβ-phase based upon PET imaging data sets we could predict the neuropathological phase of Aβ-deposition in 72.16% (70/97 cases) of the cases in our sample. When allowing a range of ± 1 Aβ-phase the Aβ-phase of 96.91% (94/97 cases) of our cases was predicted correctly based upon the PET-Aβ phase estimate. This means in this one-on-one comparison between the pathologically determined Aβ-phases and the PET-Aβ phase estimates that an incorrect prediction of the underlying neuropathological Aβ-phase was in 89% of the incorrect classified cases (24/27) in the range of the neighboring higher or lower Aβ-phase.

Discordances: Regarding the detection of Aβ-phases 1 or 2 it is noteworthy to mention that already 3/14 cases with Aβ-phases 1 or 2 were considered as amyloid positive exhibiting the criteria of PET-Aβ phase estimates 1. Here, one case showed increased cortical SUVRs in all 5 regions, one in 4/5 regions and one in 3/5 regions passing the 0.5 SUVR threshold. In two of these cases with regions with SUVR < 0.5, these PET-negative regions were the anterior cingulate gyrus (1 case), the parietal cortex (1 case) and the lateral temporal cortex (1 case). One phase 3 case did not pass the cortical but only the SUVRcaud threshold whereas the other phase 3 cases showed a heterogeneous distribution of the positive neo- and allocortical regions varying from 1–5 regions passing the cortical SUVR threshold without a distinct predilection site (number of cases with SUVR ≥ 0.5/all Aβ-phase 3 cases: anterior cingulate gyrus: 4/8; frontal cortex: 5/8; parietal cortex: 4/8; lateral temporal cortex: 4/8; posterior cingulate gyrus and precuneus: 6/8). In Aβ-phase 4 four cases showed regional SUVRs under the threshold of amyloid positivity. There were no regions spared of being negative in at least one of these four cases. Two of these cases had Braak NFT-stage I, one Braak NFT-stage II, and one Braak NFT-stage V. From the Aβ-phase 5 cases only one case showed cortical SUVRs lower than 0.5. This was seen in the anterior cingulate gyrus and in the frontal cortex. This case had Braak NFT-stage IV.