Received: 30 September 2020 Revised: 21 January 2021 Accepted: 31 January 2021 DOI: 10.1002/alz.12317 RESEARCH ARTICLE AMYQ: An index to standardize quantitative amyloid load across PET tracers Jordi Pegueroles1,3 Victor Montal1,3 Alexandre Bejanin1,3 Eduard Vilaplana1,3 Mateus Aranha1,3 Miguel Angel Santos-Santos1,3 Daniel Alcolea1,3 Ignasi Carrió2 Valle Camacho2 Rafael Blesa1,3 Alberto Lleó1,3 Juan Fortea1,3 for the Alzheimer Disease Neuroimaging Initiative* Australian Imaging, Biomarkers and Lifestyle Research Group# 1 Sant Pau Memory Unit, Department of Neurology, Hospital de la Santa Creu i Sant Abstract Pau, Biomedical Research Institute Sant Pau, Introduction: Positron emission tomography (PET) amyloid quantification methods Universitat Autònoma de Barcelona, Barcelona, Spain require magnetic resonance imaging (MRI) for spatial registration and a priori refer- 2 Department of Nuclear Medicine, Hospital de ence region to scale the images. Furthermore, different tracers have distinct thresh- la Santa Creu i Sant Pau, Universitat Autònoma olds for positivity. We propose the AMYQ index, a new measure of amyloid burden, to de Barcelona, Barcelona, Spain 3 Centro de Investigación Biomédica en Red overcome these limitations. sobre Enfermedades Neurodegenerativas Methods: We selected 18F-amyloid scans from ADNI and Australian Imaging, (CIBERNED), Madrid, Spain Biomarker & Lifestyle Flagship Study of Ageing (AIBL) with the corresponding T1-MRI. Correspondence A subset also had neuropathological data. PET images were normalized, and the AMYQ Juan Fortea Ormaechea, Memory Unit, Depart- ment of Neurology, Hospital of Sant Pau. Sant was calculated based on an adaptive template. We compared AMYQ with the Centiloid Antoni María Claret, 167. 08025. Barcelona. scale on clinical and neuropathological diagnostic performance. Spain. Email: jfortea@santpau.cat Results: AMYQ was related with amyloid neuropathological burden and had excellent * diagnostic performance to discriminate controls from patients with Alzheimer’s dis- Data used in preparation of this article were obtained from the Alzheimer’s Dis- ease (AD) (area under the curve [AUC] = 0.86). AMYQ had a high agreement with the ease Neuroimaging Initiative (ADNI) database Centiloid scale (intraclass correlation coefficient [ICC] = 0.88) and AUC between 0.94 (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and and 0.99 to discriminate PET positivity when using different Centiloid cutoffs. implementation of ADNI and/or provided data Discussion: AMYQ is a new MRI-independent index for standardizing and quantifying but did not participate in the analysis or writ- ing of this report. A complete listing of ADNI amyloid load across tracers. investigators can be found at: http://adni.loni. usc.edu/wp,content/uploads/how_to_apply/ KEYWORDS ADNI_Acknowledgement_List.pdf Alzheimer’s disease, amyloid burden standardization, amyloid pet, amyloid pet quantification, # https://aibl.csiro.au/about/ centiloid, neuropathology aibl-research-team 1 INTRODUCTION the brain.3–5 Several radiotracers have been developed for this pur- pose. 11 C-Pittsburgh compound B (PiB), the first tracer, has been used Amyloid beta (Aβ) is one of the pathophysiological markers of widely since its inception.5 However, due to the short half-life of car- Alzheimer’s disease (AD).1,2 Amyloid imaging with positron emission bon 11, several fluorine-18 (18 F)–labeled radioligands were developed: tomography (PET) enables the in vivo evaluation of Aβ deposition in 18 F-AV45, 18 F-Florbetaben, or 18 F-Flutemetamol. These PET tracers This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. © 2021 The Authors. Alzheimer’s & Dementia published by Wiley Periodicals LLC on behalf of Alzheimer’s Association Alzheimer’s Dement. 2021;17:1499–1508. wileyonlinelibrary.com/journal/alz 1499 1500 PEGUEROLES ET AL . have proved to be a reliable alternative with wider application, and have therefore been accepted by the US Food and Drug Administration RESEARCH IN CONTEXT (FDA) and the European Medicines agency (EMA).4,6,7 1. Systematic review: The literature was reviewed using Quantification of amyloid deposition in the brain is generally per- PubMed to identify publications on methodologies of formed using a standardized uptake value ratio (SUVr).8 This method- quantification of amyloid positron emission tomography ology is used widely in both cross-sectional and longitudinal amyloid (PET) imaging. PET studies.9–12 However, it has some requirements and presents sev- 2. Interpretation: This work proposes a new index that eral caveats. First, this procedure depends on two regions of interest, overcomes some of the limitations and practical difficul- one target cortical region in which to quantify amyloid and one refer- ties that have limited the implementation of amyloid PET ence region that should not be susceptible to amyloid pathology, used quantification methods in clinical practice. AMYQ is con- to intensity scale the PET images.13 The choice of the reference region sistent across tracers and does not require MRI or the is a source of variability, yielding different SUVr depending on which is definition of a priori reference and cortical regions of used.13,14 Second, due to the low anatomical information of the amy- interest. These features would facilitate its use in the clin- loid PET images, the SUVr methodology typically requires an individ- ical practice. ual 3D magnetic resonance image (MRI) for spatial normalization and 3. Future directions: Standardization of the quantification segmentation.15 Finally, the different amyloid tracers are not directly of amyloid burden across PET tracers is essential in comparable and have different thresholds for amyloid positivity.15,16 Alzheimer’s disease (AD) studies. The proposed method- To overcome these limitations, the Centiloid Working Group was ology should be replicated in other amyloid PET samples created with the aim of scaling all non-standard methodologies of amy- and longitudinal sensitivity should be assessed in order to loid PET quantification into a standard Centiloid scale.17 This scale has establish the validity of this measure. shown good results across the different amyloid tracers.16,18–22 The Centiloid scale has been tested against the neuropathological results to determine amyloid positivity thresholds.19,23,24 However, the trans- formation of the SUVr values into Centiloid requires a structural MRI, a reference region, and a set of linear transformations. T1-weighted MRI obtained <1 year apart from the PET acquisition. Given that structural MRI is not systematically available, and that More details about acquisition and pre-processing steps on PET and MRI motion artifacts could influence the accuracy of region of inter- MRI data can be found at adni.loni.usc.edu and aibl.csiro.au. est (ROI) definition,25 MRI-independent approaches have been devel- oped recently to quantify amyloid deposition based on a spatial 2.2 Neuropathology assessment normalization process that does not require MRI.26,27 Some authors Twenty-six subjects with 18 F-AV45 PET had also a neuropathologi- have proposed automated adaptive template methods, where a com- bination of different images is used to create a single template to mit- cal assessment in ADNI (the mean interval between PET acquisition igate the bias when a mean-single template is used.28–33 However, and death was 0.93 + -0.26 years). Pathological data included several the methodologies developed thus far do not obtain standard amyloid amyloid quantification scales: the Consortium to Establish a Registry measures across different tracers. for Alzheimer’s Disease (CERAD) score,34 the Thal phase pathologic Our objectives were to develop a methodology that would be able: criteria,35 and the Alzheimer’s disease neuropathic change (ADNC).36 (1) to normalize amyloid PET data without the use of MRI, and (2) to The CERAD score reflects the amyloid plaque density, particularly neu- quantify amyloid deposition in a reproducible manner across tracers ritic plaques in selected cortical areas, whereas the Thal phases reflect (AMYQ), that does not require the use of a priori reference regions, the topography of the amyloid plaques. This last assessment does not or transformations between tracers. We assessed the relationship of distinguish between compact and diffuse amyloid deposits. The ADNC AMYQ with standard neuropathological scales, and we compared it to is a composite score that uses the CERAD score and the Thal and Braak the commonly used Centiloid scale. neurofibrillary tangle stages to classify the ADNCs. More information can be found at adni.loni.usc.edu. 2 METHODS 2.3 Image processing with MRI 2.1 Participants All PET images were co-registered to their corresponding MRI using We selected a total of 751 subjects with available 18 F-AV45 PET scans a rigid body registration and warped into the Montreal Neurological and 225 subjects with available 18 F-Florbetaben PET scans from the Institute (MNI) space using an MRI-based affine transformation fol- Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort and 198 lowed by a non-linear registration using the Advanced Normalization subjects from the Australian Imaging, Biomarker & Lifestyle Flagship Tools (ANTs) software.37 SUVrs were calculated from the spatially Study of Ageing (AIBL) data set with available 18 F-Flutemetamol PET normalized images using the defined cortical volume of interest as a scans. Participants were cognitive normal, had mild cognitive impair- composite and the whole cerebellum plus brainstem as the reference ment (MCI), or an AD dementia diagnosis. All participants had a 3T region.14,17 Finally, SUVrs for the different tracers were transformed PEGUEROLES ET AL . 1501 F I G U R E 1 (A) Flow-chart of the template creation. (B) Adaptive template pipeline to normalize PET scans to the MNI space. (1) Affine transformation from native space to the MNI with the first component of the template. (2) Entering in the normalization optimization process. (3) and (4) Optimize the β0 and β1 to create an adaptive template that minimize the error with the individual PET. (5) Non-linear transformation to refine the normalization with the template. Steps 3, 4, and 5 are repeated until betas do not change into the Centiloid scale (Supplementary Material). Of note, we also erated a subject-specific adaptive template optimizing the two betas tested other volumes of interest provided by the Centiloid working of the two principal components of the PCA template (β0 for the non- group (http://www.gaain.org/Centiloid-project; Supplementary specific component, and β1 for the cortical component) by maximizing Material). the global correlation between the voxel intensities of the spatially nor- malized amyloid PET image and the adaptive template. Therefore, 2.4 Template creation PETMNI-normalized = 𝛽0 ∗ PC0 + 𝛽1 ∗ PC1 To normalize the amyloid PET scans to the MNI regardless of MRI, we created a PET template in the MNI space. For each tracer, we iteratively We then repeated the normalization step using the generated adap- and randomly selected 50 subsamples of MNI-normalized and scaled tive template using non-linear transformation from ANTs. Then we PET images of 15 controls, 10 MCI, and 15 AD subjects. For each sub- recomputed the betas to generate a new adaptive template with which sample, we applied a principal component analysis (PCA) to obtain the to again normalize the images. This process was iteratively conducted main components. The final PCA template for each tracer was created until the betas were optimized and did not change from the previous as the mean of the 50 permutations of the previously obtained com- normalization step. The algorithm finally created a unique amyloid-PET ponents. For all three tracers, the first two components of the PCA template for each subject using the optimal estimated betas for each of explained >90% of the variance of the data (Figure 1A). The visual the two components of the template. inspection of each component identified the first component (PC0) as We assessed the normalization procedure of each tracer with the the non-specific binding of the amyloid tracer (ie, mainly white matter corresponding tracer-specific template (ie, data-driven specific and binding), whereas the second one (PC1) represented the specific cor- unspecific binding), but we also used the different tracer-specific tical binding. More details about template creation are provided in the templates in the other tracer subsamples. We compared the global Supplementary Material. Pearson’s correlation between templates, and they showed high correspondence between their specific and unspecific binding com- 2.5 Image processing with adaptive template and ponents (Supplementary Material). The 18 F-Flutemetamol template AMYQ index outperformed the two other templates (ie, higher agreement with the Centiloid scale) in all subsamples. Therefore, we present the All 1174 individual amyloid PET scans were normalized to the MNI results using the PCA template created using 18 F-Flutemetamol for space using a linear combination of the two components of the PCA all the analyses. For completeness, the results using the two other two template. Figure 1B shows the complete process. The native PET image templates are shown in the Supplementary Material. was initially normalized with an affine registration to the PC0 using The AMYQ index was determined as the β1/β0 ratio. We used the mutual information as similarity measure. To refine the normalization first component (or PC0) as a data-driven reference region to scale the to MNI space, we performed an optimization process. First, we gen- cortical component and the second (or PC1) as the data-driven region 1502 PEGUEROLES ET AL . in which to measure amyloid deposition. The ratio was then trans- 3 RESULTS formed into a 100-point scale using the same approach as that used by Klunk et al.17 Taking advantage of the 18 F amyloid data set from Global 3.1 Participants Alzheimer’s Association Interactive Network (GAAIN),20–22 we calcu- lated the mean β1/β0 ratio in the young controls and in the AD patients. PET images were successfully normalized with the adaptive template The β1/β0 ratio of the young controls was mean centered to 0 AMYQ in 1148 subjects (of 1174) in contrast to the 1106 scans (of 1174), and to 100 AMYQ in AD patients. Thus the AMYQ for each individual which where correctly warped to the MNI with the MRI-based normal- was defined as: ization. Of note, 12 subjects failed with both methodologies; 56 failed only with the MRI-based normalization, whereas only 14 failed partic- 𝛽1∕𝛽0ind − 𝛽1∕𝛽0YC AMYQ = 100 ∗ ularly with the adaptive template approach. Details on the normaliza- 𝛽1∕𝛽0AD − 𝛽1∕𝛽0YC tion failure are provided in the Supplementary Material. Therefore, the As with the Centiloid scale, higher values of AMYQ represent high normalization process was successful with both methodologies in 1092 amyloid burden (ie, AD subjects mean scaled to 100 AMYQ), whereas subjects. Table 1 summarizes the demographics of these 1092 partici- values around zero represent low amyloid burden (ie, young controls pants included in the subsequent analyses: 434 (39.7%) were cognitive mean scaled to zero). AMYQ, as the Centiloid scale, can exceed these normal, 505 (46.3%) had MCI, and 153 (14.0%) had an AD dementia limits. diagnosis. 2.6 Statistical analysis 3.2 Assessing quality control of normalization procedure Analyses were performed with the R statistical software (v 3.6.3; https: //www.r-project.org). To assess the quality of the spatial normaliza- The estimated similarity between the spatially normalized PET images tion of the adaptive template methodology, we computed the similarity warped using MRI and the PCA template was accurate across trac- structural image similarity (SSIM)38 index between the two normalized ers (SSIMAV45 = 0.85 + −0.06, SSIMFlorbetaben = 0.85 + −0.05, and images. SSIMFlutemetamol = 0.91 + −0.04. To assess the relationship of AMYQ with post-mortem amyloid deposition, we computed the Spearman correlation between the four neuropathologic scales and both the AMYQ and the Centiloid scores. Kruskal-Wallis and Dunn tests were performed to test differences 3.3 AMYQ correlates with the different with the neuropathologic assessments for both measures. In addition, neuropathological scales paired Wilcoxon signed-rank tests were performed within each score of the neuropathological scales to test differences between AMYQ and Figure 2 shows the distribution of AMYQ and Centiloid scores Centiloid measures. across the different post-mortem amyloid quantifications. Both mea- To derive AMYQ thresholds, receiver-operating characteristic sures increased significantly with higher scores in the four neu- (ROC) curves were calculated to assess the ability of AMYQ and Cen- ropathologic scales (ρAMYQ = 0.474, ρCentiloid = 0.562; ρAMYQ = 0.443, tiloid to distinguish between controls and AD dementia patients in ρCentiloid = 0.551; ρAMYQ = 0.392, ρCentiloid = 0.546; and ρAMYQ = 0.610, the three subsamples separately, and in the combined amyloid sample. ρCentiloid = 0.660 for Thal phases, ADNC, and CERAD neuritic and dif- Youden index was used to establish the optimal thresholds. In addition, fuse, respectively). the DeLong test was used to compare the performance of both amy- AMYQ and the Centiloid scale showed similar results in all the loid measures to distinguish controls from AD. Then, between-clinical amyloid neuropathological stages. There were no statistically signifi- group effect sizes (Hedges g) were calculated for both AMYQ and Cen- cant differences between the Thal phase 0 and phase 1, “none” ver- tiloid measures. sus “low” ADNC, or “none” versus “sparse” amyloid pathology using To assess the agreement between AMYQ and Centiloid, we com- CERAD. Both AMYQ and Centiloid showed differences with the other puted the ICC between both metrics for each tracer separately and Thal phases, when compared with Thal phase 0 or “none” pathology in the combined amyloid sample. In addition, Bland-Altman plots were (P = .025 and P = .022 vs phase 4; P = .025 and P = .013 vs phase used to compare both metrics. Finally, ROC curves were conducted to 5); with the ADNC scores (P = .012 and P = .010 vs “intermediate”; assess the power of AMYQ to discriminate between positive and nega- P = .024 and P = .015 vs “high”); and for both CERAD neuritic “fre- tive amyloid-PET scans. To dichotomize the data into amyloid-positive quent” (P = .029 and P = .005), and diffuse “frequent” plaques (P = .020 and amyloid-negative individuals, we used a Centiloid cutoff of 12.2, and P = .014 for AMYQ and Centiloid, respectively). The Wilcoxon which has been reported to identify Aβ-detectable Thal phases and to ranked test did not reveal significant differences between AMYQ and identify “moderate/frequent” CERAD amyloid burden.24 Other thresh- the Centilod scales in any of the scores of the four neuropathological olds were also used and are reported in the Supplementary Material. scales. PEGUEROLES ET AL . 1503 [66.90, 78.30] [70.80, 81.55] [68.00, 76.00] [69.00, 78.50] [71.00, 79.00] 143.94] 115.85] TA B L E 1 Demographics for the overall sample and the three subsamples divided by diagnosis (Control, MCI, mild cognitive impairment, and AD, Alzheimer’s disease dementia). Median and 9 ( 52.9) [66.09, [−5.39, 36.30] [11.01, 94.25] [74.71, 104.30 75.00 90.99 AD 17 102.72] 20 ( 51.3) [11.65, 74.00 65.50 65.45 MCI 39 Flutemetamol [4.74, 30.60] 57 ( 40.7) Control 72.00 12.80 7.23 140 109.35] 122.91] 9 ( 60.0) [−6.86, 65.29] [54.71, [65.47, 103.29 73.60 87.39 AD 15 38 ( 62.3) 55.36] [−25.83, 17.85] [−26.09, 74.00 12.76 6.43 MCI 61 F I G U R E 2 Boxplots showing the distribution of the AMYQ and [67.50, 76.00] [66.20, 77.00] [71.00, 79.70] [68.53, 77.35] [66.00, 76.20] [71.00, 79.50] [67.00, 74.53] [−6.63, 15.74] Florbetaben Centiloid measures through the different neuropathological scales 56 ( 38.9) Thal phases, AD neuropathological change (ADNC), CERAD neuritic Control −3.79 69.70 plaques, and CERAD diffuse plaques. *Indicates significant differences 0.99 144 between scores of the different scales with both measures (AMYQ and Centiloid) at P < .05 121.74] 107.77] 68 ( 56.2) [66.35, [60.55, 3.4 AMYQ distinguishes between clinical groups 75.30 94.28 88.29 121 AD The ROC curve using AMYQ for the discrimination between clinical [−7.51, 32.25] [0.27, 83.20] 225 ( 55.6) groups did not show significant differences between AMYQ and the 79.04] [−10.12, Centiloid scales. Both showed a good performance across the three 71.50 28.28 22.13 MCI 405 tracers for the discrimination between controls and AD dementia patients. The effect sizes for between-group comparisons were simi- lar across methods and tracers (Table 2). Both AMYQ and the Centiloid 73 ( 48.7) 27.83] [−28.43, Control −6.40 scale had high accuracy in the discrimination between controls and AD 72.80 AV45 interquartile range (IQR) are shown for each continuous variable 6.70 150 dementia patients in the combined amyloid sample (AUCAMYQ = 0.84, AUCCentiloid = 0.86, P > .05). The threshold for positivity derived from 122.18] 111.16] the Youden index were 48.9 for AMYQ and 41.9 in the Centiloid scale. 86 ( 56.2) [−3.98, 26.77] [−0.35, 83.20] [65.37, [61.86, 75.00 94.67 90.29 153 AD 3.5 Good agreement between the AMYQ and 283 ( 56.0) Centiloid scales 79.04] [−10.17, 72.00 26.53 22.38 MCI 505 The AMYQ index had a high agreement with the Centiloid for the different PET tracers separately (Figure 3; ICCAV45 = 0.88, Gender = M 186 ( 42.9) ICCFlorbetaben = 0.88, and ICCFlutemetamol = 0.89), and in the combined 27.30] [−20.78, Control Overall sample (ICC = 0.88). The Bland-Altman plot showed a bias of vari- 71.60 6.68 1.06 434 ances across the measurement range (P < .001). When stratifying by amyloid positivity (as determined by a 12.2 Centiloid), the negative Age (years) (median (median (median Centiloid [IQR]) [IQR]) [IQR]) subsample showed a wider heterogeneity within the range of values AMYQ (%) (r = −0.71; P < .001), whereas no bias was found in the positive sub- N sample (r = 0.08; P > .05). 1504 PEGUEROLES ET AL . F I G U R E 3 In the upper row, scatterplots showing the agreement between the AMYQ and Centiloid scales in the three tracer subsamples (from left to right: AV45, Florbetaben, and Flutemetamol) and the overall sample (on the right). In the lower row, Bland-Altman plots showing the difference between Centiloid and AMYQ by tracer and in the overall sample The discriminatory power of AMYQ to detect amyloid positivity AMYQ is based on a synthetic amyloid template generated using was very high in the combined sample and in the three samples sep- PCA. The topographical pattern of this template is in agreement with arately (AUCCombined = 0.94, AUCAV45 = 0.94, AUCFlorbetaben = 0.97, previous tracer-specific works where one component represents the and AUCFlutemetamol = 0.89). An AMYQ threshold of 15.6 discriminated unspecific binding, mostly comprising white matter regions, and a between amyloid positivity groups, with 92% specificity and 84% sen- second component that shows the neocortical uptake.32,33 Applying sitivity in the combined sample. The aforementioned PET-positivity a linear combination of the specific and non-specific components by threshold of 15.6 AMYQ had a sensitivity of 86% and a specificity of iteratively optimizing the weights of the two components of the PCA 68% in the discrimination between controls and AD dementia patients template, we created a unique adaptive template for each amyloid in the overall sample. Of note, other established Centiloid thresholds scan that will work as standard space to warp the individual image. to determine PET positivity23,39 yielded AUCs of AMYQ between 0.97 By computing the ratio between these two weights, we obtained and 0.99 (Supplementary Material). a measure of amyloid burden that accounts for the differences of intensity scale between PET acquisitions. One of the main advantages of this methodology is that it does not depend upon predefined flat 4 DISCUSSION (non-weighted) regions of cortical load or reference region to scale the PET. It is important to note that the weighted data-driven regions The AMYQ index is a new metric to quantify brain amyloid load that derived from the PCA analysis allow the assigning of more weight to does not require an MRI. AMYQ can be immediately computed in key spatial regions related with the amyloid uptake pattern (eg, for the nuclear medicine departments after the amyloid PET acquisition, the precuneus or medial frontal regions). Therefore, this approach providing a quantitative assessment that could help in clinical prac- removes the variability associated with the selection of distinct regions tice, which currently usually relies only on a visual read. Furthermore, (both the reference region and the cortical region), and facilitates the AMYQ is interchangeable across tracers, and thus provides a standard- harmonization between methods. To ensure the robustness of our ized measure for amyloid quantification that could facilitate the com- results, we created a data-driven template for each tracer and com- parison of studies in which different tracers are used and in clinical pared the performance of each template onto all subsamples. All had practice. good performances when applied to the other tracers, which is not PEGUEROLES ET AL . 1505 surprising given that all tracers have been related with Aβ neu- ropathology and share both the specific and unspecific binding (as AUC 0.89 0.75 0.71 shown in Supplementary Material). The 18 F-Flutemetamol template had a higher non-specific binding in the white matter than the other two tracers, as described previously.11,40–42 In this respect, a recent Hedges g study using fluorodeoxyglucose (FDG)-PET, reported a remarkable AMYQ 2.00 1.06 0.74 bias when intensity normalization methods were used (instead of data- driven methods), and recommended the use of reference regions with large volume and good stability.43 Accordingly, the data-driven refer- AUC ence regions derived from the 18 F-Flutemetamol subsample proved 0.89 0.71 0.71 to be superior to those derived from the other two subsamples. It is important to emphasize that AMYQ, which can be assimilated to a data- driven SUVr, is based in larger and more stable cortical and reference Flutemetamol regions than those commonly used in a priori reference regions.11–13,15 Centiloid Hedges g AMYQ was related to post-mortem amyloid pathology, both with 2.33 1.10 0.77 CERAD the Thal phases, thus indicating that AMYQ is related with not only the abundance of neuritic plaques but also with the spatial- temporal distribution of amyloid deposition. The correlation between post-mortem amyloid pathology and the AMYQ and Centiloid scales AUC 0.87 0.58 0.82 was similar, and in agreement with previous Centiloid reports.21,23,24 These data highlight the association between the amyloid PET quanti- tative measure and the severity of the neuropathology staging, which Hedges g AMYQ supports the concept of monitoring AD disease progression through 1.92 0.30 1.27 amyloid PET. Effect sizes (Hedges g) and AUC between diagnoses for both Centiloid and AMYQ across tracers The discriminatory power of AMYQ when differentiating between AD dementia patients and controls showed excellent accuracy, equiv- alent to that of the Centiloid scale. For all three tracers, the AMYQ AUC 0.56 0.82 0.75 and the Centiloid scale index showed similar effect sizes and AUC to discriminate between the different diagnoses. In previous studies, Florbetaben Whittington et al.33 have shown a very good diagnostic performance Centiloid Hedges g of their amyloid load measure in 18 F-AV45 samples. Their measure 0.41 1.96 1.04 showed a higher sensitivity than the commonly used SUVr in all comparisons.33,44 Of note, in the 18 F-AV45 subsample of our work, AMYQ showed effect sizes similar to those reported by the Whitting- ton measure. The agreement between the AMYQ and Centiloid values AUC 0.67 0.86 0.74 was excellent, both when assessed for the three tracers separately and when assessed in the combined sample. This yielded similar thresholds for AMYQ and the Centiloid scale when comparing controls and AD Hedges g dementia patients. Of note, the thresholds were similar to those AMYQ 0.60 0.86 1.58 reported in other studies using the Centiloid scale.45 Of note, these thresholds were also similar to those reported when assessing agree- ment with CSF p-Tau/Aβ42 and t-Tau/Aβ42 ratios in the ADNI cohort.39 Despite the high correspondence between measures, we observed dis- AUC 0.65 0.73 0.84 sociation of values between the AMYQ and Centiloid in the subsample with low amyloid burden, but not in the amyloid-positive range. The ability of AMYQ to detect amyloid positivity using different Centiloid Hedges g Centiloid cutoffs was also very high, with AUCs systematically above AV45 0.52 0.87 1.51 0.94 in the overall sample and in the three subsamples separately. Using the Centiloid cutoff of 12.2 as the gold standard (this cutoff has been proposed to differentiate none/sparse from moderate/high CN—MCI MCI—AD TA B L E 2 CN—AD CERAD scores and to identify Aβ-detectable Thal phase),24 our analy- ses identified a very similar cutoff of 15.6 for AMYQ. When this thresh- old was used to differentiate patients with clinical diagnosis of AD 1506 PEGUEROLES ET AL . dementia from controls, it proved very sensitive but with lower speci- which was in addition restricted to the 18 F-AV45 subsample. Further ficity for both scales; this low specificity is probably due to the high per- analyses with larger sample size with neuropathological data should be centage of controls with amyloid pathology (40.8%). Higher Centiloid assessed. Finally, Centiloid is widely used in research settings. While thresholds, between 23.5 and 26.0, might be better suited to identify AMYQ could be quickly implemented in nuclear medicine departments, intermediate/higher ADNC24 and, importantly, to correlate with visual its use in research settings in which MRIs are usually acquired simulta- reads.23,39 neously will depend on its validation and proof of potential additional Automated adaptive template methods have been used typically to advantages, such as to measure longitudinal amyloid load or to capture normalize amyloid PET images, but not to replace the use of refer- early amyloid deposition. ence region to scale the intensity of the image. One of the most widely In summary, our study provides a new index of global amyloid load used methodologies to normalize amyloid PET without MRI is the that does not require a structural MRI and is independent from a priori CapAIBL,18,19,28 which reported excellent normalization results. Oth- reference regions. AMYQ could be used directly in clinical practice to ers have also reported an improved normalization using their respec- quantify amyloid load consistently across amyloid tracers. tive adaptive templates in 18 F-Flutemetamol scans32,46 or 11 C-PiB.26 The PCA-template normalization showed a good similarity with the ACKNOWLEDGMENTS MRI normalization, and showed very good agreement between SUVr in Data collection and sharing for this project were funded by the the three tracer subsamples and similar performance to those methods Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Insti- reported in the literature. tutes of Health grant U01 AG024904) and DOD ADNI (Department of The advantage of AMYQ over the aforementioned automated adap- Defense award W81XWH-12-2-0012). ADNI is funded by the National tive template methods is that it provides a quantitative assessment Institute on Aging, the National Institute of Biomedical Imaging and of the amyloid load that does not require MRI and is interchange- Bioengineering, and through generous contributions from the fol- able across tracers. Thus, AMYQ can be immediately computed in the lowing: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery nuclear medicine departments after the amyloid PET acquisition, and Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers be used to complement the visual read. Visual reads have been clas- Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharma- sically implemented as a binary assessment. However, the number of ceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La borderline cases is non-negligible and, considering that the subthresh- Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE old Aβ positivity could be indicative of faster AD pathology accumu- Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research lation, the dichotomization may overlook those at-risk individuals.47 & Development, LLC.; Johnson & Johnson Pharmaceutical Research Thus, AMYQ could aid in clinical practice as an immediate tool for PET & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso quantification readily available at the nuclear medicine department. By Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; using AMYQ, the practical difficulties imposed by the need of MRI usu- Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; ally acquired in a different visit in a different department or center, Servier; Takeda Pharmaceutical Company; and Transition Therapeu- which have contributed to the limited implementation of quantification tics. The Canadian Institutes of Health Research is providing funds methods in clinical practice,48 are avoided. to support ADNI clinical sites in Canada. Private sector contributions Some limitations should be considered. First, this work proves that are facilitated by the Foundation for the National Institutes of Health the AMYQ index is consistent using different templates randomly gen- (www.fnih.org). The grantee organization is the Northern California erated from a wide heterogeneous sample, but does not discard bias on Institute for Research and Education, and the study is coordinated by homogenous samples. To determine whether the use of the “universal” the Alzheimer’s Therapeutic Research Institute at the University of Flutemetamol template is preferable over a single tracer template, we Southern California. ADNI data are disseminated by the Laboratory would require calculating AMYQ indices in a sample submitted with dif- for Neuro Imaging at the University of Southern California. ferent radiotracers and PET scanners. Such samples were not available Core funding for the AIBL study was provided in part by the study in this study. In addition, this work does not definitely solve whether partners (Australian Commonwealth Scientific and Industrial Research the “universal” template should always be used instead of a specific Organization [CSIRO], Edith Cowan University [ECU], Mental Health template created for a particular study. The later approach might prove Research Institute [MHRI], Alzheimer’s Australia [AA], National Age- to be more accurate within a specific cohort, but it could compromise ing Research Institute [NARI], Austin Health, CogState Ltd., Holly- the comparability of the indices calculated. Second, the relationship wood Private Hospital, Sir Charles Gardner Hospital). The study also between AMYQ and Centiloid in the amyloid-negative subjects should received support from the National Health and Medical Research be further explored against a pathological gold standard. Moreover, Council (NHMRC) and the Dementia Collaborative Research Centres this index has not been assessed in cohorts with other amyloid trac- program (DCRC2), as well as ongoing funding from the Science and ers such as 18 F-NAV-469 or 11 C-PiB and it should be tested in other Industry Endowment Fund (SIEF). replication cohorts to validate the reported thresholds, and evaluated In addition, this study was supported by the Fondo de Investiga- longitudinally to assess its sensitivity to change in time as this is a cross- ciones Sanitario (FIS), Instituto de Salud Carlos III (PI14/01126 and sectional study. Despite the big sample size in some of the analysis, PI17/01019 to JF, PI13/01532 and PI16/01825 to RB, PI18/00435 most of the sample had absence of neuropathological confirmation, and INT19/00016 to DA, and PI14/1561 and PI17/01896 to AL) and PEGUEROLES ET AL . 1507 the CIBERNED program (Program 1, Alzheimer Disease to Alberto 5. Klunk WE, Engler H, Nordberg A, et al. Imaging brain amyloid Lleó and SIGNAL study, www.signalstudy.es), partly jointly funded by in Alzheimer’s disease with Pittsburgh Compound-B. Ann Neurol. 2004;55(3):306-319. Fondo Europeo de Desarrollo Regional, Unión Europea, Una manera 6. Nelissen N, Van Laere K, Thurfjell L, et al. Phase 1 Study of the Pitts- de hacer Europa. This work was also supported by the National Insti- burgh compound B derivative 18F-Flutemetamol in healthy volun- tutes of Health (NIA grants 1R01AG056850 - 01A1; R21AG056974 teers and patients with probable Alzheimer’s disease. J Nucl Med. and R01AG061566 to JF), Departament de Salut de la Generali- 2009;50(8):1251-1259. 7. Rowe CC, Ackerman U, Browne W, et al. Imaging of amyloid β in tat de Catalunya, Pla Estratègic de Recerca i Innovació en Salut Alzheimer’s disease with 18F-BAY94-9172, a novel PET tracer: proof (SLT002/16/00408 to A.L), Fundació La Marató de TV3 (20141210 to of mechanism. Lancet Neurol. 2008;7(2):129-135. JF and 044412 to RB); VM is supported by Fondo de Investigaciones 8. Cselényi Z, Farde L. Quantification of blood flow-dependent compo- Sanitario (FI18/00275). This work was also supported by Generali- nent in estimates of beta-amyloid load obtained using quasi-steady- state standardized uptake value ratio. J Cereb Blood Flow Metab. tat de Catalunya (SLT006/17/00119 to JF, SLT006/17/95 to EV, and 2015;35(9):1485–1493. http://doi.org/10.1038/jcbfm.2015.66. SLT006/17/00125 to DA) and a grant from the Fundació Bancaria La 9. Aghakhanyan G, Vergallo A, Gennaro M, et al. The precuneus-a witness Caixa to RB. for excessive aβ gathering in alzheimer’s disease pathology. Neurode- gener Dis. 2019:302-309.Published online. COMPETING INTERESTS 10. Bullich S, Villemagne VL, Catafau AM, et al. Optimal reference region to measure longituDinal amyloid-b change with 18F-Florbetaben PET. AL has served at the scientific advisory boards of Biogen, Fujirebio J Nucl Med. 2017;58(8):1300-1306. Europe, Eli Lilly, Novartis, Roche Diagnostics, and Nutricia. JF has 11. Mountz JM, Laymon CM, Cohen AD, et al. Comparison of qual- served at the scientific advisory boards of AC Inmune and has provided itative and quantitative imaging characteristics of [11C]PiB and consultancies to Novartis and Merck as part of the Medavante adju- [18F]flutemetamol in normal control and Alzheimer’s subjects. Neu- roImage Clin. 2015;9:592-598. dication team in the Merck MK-8931 and Novartis CAPI015A2201J 12. Shokouhi S, Mckay JW, Baker SL, et al. Reference tissue normalization and CCNP520A2202J trials. DA participated in advisory boards from in longitudinal 18F-florbetapir positron emission tomography of late Fujirebio-Europe and Roche Diagnostics and received speaker hono- mild cognitive impairment. Alzheimers Res Ther. 2016;8(1):2. raria from Fujirebio-Europe, Nutricia, and from Krka Farmacéutica S.L. 13. Chiao P, Bedell BJ, Avants B, et al. Impact of reference and target region selection on amyloid PET SUV ratios in the phase 1B PRIME DA, JF, and AL declare a filed patent application (WO2019175379 study of aducanumab. J Nucl Med. 2019;60(1):100-106. A1 Markers of synaptopathy in neurodegenerative disease). All other 14. Su Y, Flores S, Hornbeck RC, et al. Utilizing the Centiloid scale in authors report no competing interests. cross-sectional and longitudinal PiB PET studies. NeuroImage Clin. 2018;19(April):406-416. 15. Landau SM, Thomas BA, Thurfjell L, et al. Amyloid PET imaging in ORCID Alzheimer’s disease: a comparison of three radiotracers. Eur J Nucl Med Jordi Pegueroles https://orcid.org/0000-0002-3554-2446 Mol Imaging. 2014;41(7):1398-1407. Victor Montal https://orcid.org/0000-0002-5714-9282 16. Su Y, Flores S, Wang G, et al. Comparison of Pittsburgh compound B Alexandre Bejanin https://orcid.org/0000-0002-9958-0951 and florbetapir in cross-sectional and longitudinal studies. Alzheimer’s Dement Diagnosis. Assess Dis Monit. 2019;11:180-190. Eduard Vilaplana https://orcid.org/0000-0002-1809-8435 17. Klunk WE, Koeppe RA, Price JC, et al. The Centiloid Project: stan- Mateus Aranha https://orcid.org/0000-0001-9594-292X dardizing quantitative amyloid plaque estimation by PET. Alzheimer’s Miguel Angel Santos-Santos https://orcid.org/0000-0003-1567- Dement. 2015;11(1):1-15.e4. 617X 18. Bourgeat P, Doré V, Fripp J, et al. Implementing the centiloid trans- Daniel Alcolea https://orcid.org/0000-0002-3819-3245 formation for11C-PiB and β-amyloid18F-PET tracers using CapAIBL. Neuroimage. 2018;183:387-393. Ignasi Carrió https://orcid.org/0000-0003-1135-5352 19. Doré V, Bullich S, Rowe CC, et al. Comparison of 18 F-florbetaben Valle Camacho https://orcid.org/0000-0003-0748-0847 quantification results using the standard Centiloid, MR-based, and Rafael Blesa https://orcid.org/0000-0003-4026-2884 MR-less CapAIBL ® approaches: validation against histopathology. Alberto Lleó https://orcid.org/0000-0002-2568-5478 Alzheimer’s Dement. 2019:1-10.Published online. 20. Battle MR, Pillay LC, Lowe VJ, et al. Centiloid scaling for quantifica- Juan Fortea https://orcid.org/0000-0002-1340-638X tion of brain amyloid with [18F]flutemetamol using multiple process- ing methods. EJNMMI Res. 2018;8(1):107. REFERENCES 21. Navitsky M, Joshi AD, Kennedy I, et al. Standardization of amyloid 1. Fortea J, Vilaplana E, Alcolea D, et al. Cerebrospinal fluid β-amyloid and quantitation with florbetapir standardized uptake value ratios to the phospho-tau biomarker interactions affecting brain structure in pre- Centiloid scale. Alzheimer’s Dement. 2018;14(12):1565-1571. clinical Alzheimer disease. Ann Neurol. 2014;76(2). 22. Rowe CC, Doré V, Jones G, et al. 18F-Florbetaben PET beta- 2. Mattsson N, Insel PS, Nosheny R, et al. Emerging β-Amyloid pathol- amyloid binding expressed in Centiloids. Eur J Nucl Med Mol Imaging. ogy and accelerated cortical atrophy. JAMA Neurol. 2014;71(6): 2017:2053-2059.Published online. 725. 23. Amadoru S, Doré V, McLean CA, et al. Comparison of amyloid 3. Jagust W. Is amyloid-β harmful to the brain? Insights from human imag- PET measured in Centiloid units with neuropathological findings in ing studies. Brain. 2016;139(1):23–30. http://doi.org/10.1093/brain/ Alzheimer’s disease. Alzheimer’s Res Ther. 2020;12(1):22. awv326. 24. La Joie R, Ayakta N, Seeley WW, et al. Multisite study of the relation- 4. Wong DF, Rosenberg PB, Zhou Y, et al. In vivo imaging of amyloid depo- ships between antemortem [11 C]PIB-PET Centiloid values and post- sition in Alzheimer disease using the radioligand 18F-AV-45 (flobe- mortem measures of Alzheimer’s disease neuropathology. Alzheimer’s tapir F 18). J Nucl Med. 2010;51(6):913-920. Dement. 2019;15(2):205-216. 1508 PEGUEROLES ET AL . 25. Imabayashi E, Matsuda H, Tabira T, et al. Comparison between brain CT 39. Salvadó G, Molinuevo JL, Brugulat-Serrat A, et al. Centiloid cut-off val- and MRI for voxel-based morphometry of Alzheimer’s disease. Brain ues for optimal agreement between PET and CSF core AD biomarkers. Behav. 2013;3(4):487-493. Alzheimers Res Ther. 2019;11(1):27. 26. Akamatsu G, Ikari Y, Ohnishi A, et al. Automated PET-only quantifica- 40. Lowe VJ, Lundt E, Knopman D, et al. Comparison of [18F]Flutemetamol tion of amyloid deposition with adaptive template and empirically pre- and [11C]Pittsburgh Compound-B in cognitively normal young, cog- defined ROI. Phys Med Biol. 2016;61(15):5768-5780. nitively normal elderly, and Alzheimer’s disease dementia individuals. 27. Edison P, Carter SF, Rinne JO, et al. Comparison of MRI based and NeuroImage Clin. 2017;16(July):295-302. PET template based approaches in the quantitative analysis of amyloid 41. Adamczuk K, Schaeverbeke J, Nelissen N, et al. Amyloid imag- imaging with PIB-PET. Neuroimage. 2013;70:423-433. ing in cognitively normal older adults: comparison between 18F- 28. Bourgeat P, Dore V, Fripp J, Villemagne VL, Rowe CC, Salvado O, Com- flutemetamol and 11C-Pittsburgh compound B. Eur J Nucl Med Mol putational analysis of PET by AIBL CapAIBL: a cloud-based process- Imaging. 2016;43(1):142-151. ing pipeline for the quantification of PET images [abstract]. J Nucl Med. 42. Cho SH, Choe YS, Kim YJ, et al. Head-to-Head comparison of 18f- 2015;56(suppl 3):149. florbetaben and 18f-flutemetamol in the cortical and striatal regions. J 29. Bourgeat P, Villemagne VL, Dore V, et al. Comparison of MR-less Alzheimer’s Dis. 2020:1-10.Published online. PiB SUVR quantification methods. Neurobiol Aging. 2015;36:S159- 43. López-González FJ, Silva-Rodríguez J, Paredes-Pacheco J, et al. Inten- S166. sity normalization methods in brain FDG-PET quantification. Neuroim- 30. Kang SK, Seo S, Shin SA, et al. Adaptive template generation for age. 2020;222(July). amyloid PET using a deep learning approach. Hum Brain Mapp. 44. Tanaka T, Stephenson MC, Nai YH, et al. Improved quantification of 2018;39(9):3769-3778. amyloid burden and associated biomarker cut-off points: results from 31. Lilja J, Leuzy A, Chiotis K, Savitcheva I, Sörensen J, Nordberg the first amyloid Singaporean cohort with overlapping cerebrovascu- A. Spatial normalization of [18 F]flutemetamol PET images uti- lar disease. Eur J Nucl Med Mol Imaging. 2020;47(2):319-331. lizing an adaptive principal components template. J Nucl Med. 45. Leuzy A, Chiotis K, Hasselbalch SG, et al. Pittsburgh compound B imag- 2018:jnumed.118.207811. ing and cerebrospinal fluid amyloid-β in a multicentre European mem- 32. Lundqvist R, Lilja J, Thomas BA, et al. Implementation and validation ory clinic study. Brain. 2016;139(9):2540-2553. of an adaptive template registration method for 18F-Flutemetamol 46. Lilja J, Leuzy A, Chiotis K, Savitcheva I, Sörensen J, Nordberg A. Spa- Imaging Data. J Nucl Med. 2013;54(8):1472-1478. tial normalization of 18 F-flutemetamol PET images using an adaptive 33. Whittington A, Gunn RN. Amyloid load: a more sensitive biomarker for principal-component template. J Nucl Med. 2019;60(2):285-291. amyloid imaging. J Nucl Med. 2019;60(4):536-540. 47. Bischof GN, Jacobs HIL. Subthreshold amyloid and its biological and 34. Mirra SS, Heyman A, McKeel D, et al. The consortium to establish clinical meaning: long way ahead. Neurology. 2019;93(2):72-79. a registry for Alzheimer’s disease (CERAD). Part II. Standardization 48. Mainta IC, Vargas MI, Trombella S, Frisoni GB, Unschuld PG, Garibotto of the neuropathologic assessment of Alzheimer’s disease. Neurology. V. Hybrid PET-MRI in Alzheimer’s disease research. In: methods in 1991;41(4):479-486. Molecular Biology. Humana Press Inc. 2018;1750:185-200. 35. Thal DR, Rüb U, Orantes M, Braak H. Phases of Aβ-deposition in the human brain and its relevance for the development of AD. Neurology. SUPPORTING INFORMATION 2002;58(12):1791-1800. 36. Hyman BT, Phelps CH, Beach TG, et al. National Institute on Additional supporting information may be found online in the Support- Aging-Alzheimer’s Association guidelines for the neuropathologic ing Information section at the end of the article. assessment of Alzheimer’s disease. Alzheimer’s Dement. 2012;8(1): 1-13. 37. Avants BB, Epstein CL, Grossman M, Gee JC. Symmetric diffeomor- phic image registration with cross-correlation: evaluating automated How to cite this article: Pegueroles J, Montal V, Bejanin A, labeling of elderly and neurodegenerative brain. Med Image Anal. et al. AMYQ: An index to standardize quantitative amyloid load 2008;12(1):26-41. 38. Wang Z, Bovik AC, Sheikh HR, Simoncelli EP. Image quality assess- across PET tracers. Alzheimer’s Dement. 2021;17:1499–1508. ment: from error visibility to structural similarity. IEEE Trans Image Pro- https://doi.org/10.1002/alz.12317 cess. 2004;13(4):600-612.