A multicenter comparison of [18F]flortaucipir, [18F]RO948, and [18F]MK6240 tau PET tracers to detect a common target ROI for differential diagnosis

Purpose This study aims to determine whether comparable target regions of interest (ROIs) and cut-offs can be used across [18F]flortaucipir, [18F]RO948, and [18F]MK6240 tau positron emission tomography (PET) tracers for differential diagnosis of Alzheimer’s disease (AD) dementia vs either cognitively unimpaired (CU) individuals or non-AD neurodegenerative diseases. Methods A total of 1755 participants underwent tau PET using either [18F]flortaucipir (n = 975), [18F]RO948 (n = 493), or [18F]MK6240 (n = 287). SUVR values were calculated across four theory-driven ROIs and several tracer-specific data-driven (hierarchical clustering) regions of interest (ROIs). Diagnostic performance and cut-offs for ROIs were determined using receiver operating characteristic analyses and the Youden index, respectively. Results Comparable diagnostic performance (area under the receiver operating characteristic curve [AUC]) was observed between theory- and data-driven ROIs. The theory-defined temporal meta-ROI generally performed very well for all three tracers (AUCs: 0.926–0.996). An SUVR value of approximately 1.35 was a common threshold when using this ROI. Conclusion The temporal meta-ROI can be used for differential diagnosis of dementia patients with [18F]flortaucipir, [18F]RO948, and [18F]MK6240 tau PET with high accuracy, and that using very similar cut-offs of around 1.35 SUVR. This ROI/SUVR cut-off can also be applied across tracers to define tau positivity. Supplementary Information The online version contains supplementary material available at 10.1007/s00259-021-05401-4.


Introduction
In addition to neuritic plaques composed of amyloid-β (Aβ), Alzheimer's disease (AD) is characterized by tau pathology, largely in the form of paired helical filaments (PHFs) comprising a mixture of three/four-repeat tau isoforms [1]. While tau aggregates are also present in various non-AD neurodegenerative disorders, including certain variants of frontotemporal lobar degeneration and progressive supranuclear palsy (PSP), these are structurally distinct from those observed in AD [2,3].
The development of radiotracers selective for tau aggregates for use with positron emission tomography (PET) has allowed for their visualization and quantification in vivo [4]. The most widely used tau tracer to date, [ 18 F]flortaucipir [5], has been shown to primarily detect AD-type tau aggregates [6,7]; as such, tau PET may be most valuable for differentiating AD from non-AD tauopathies and other neurodegenerative disorders. Though tau PET is a relatively recent technique, several novel tracers characterized by improved specificity and dynamic range have recently entered the field, including [ 18 F]RO948 [8] and [ 18 F]MK6240 [9,10]. While characterized by improved specificity and dynamic range, these tracers also show greater meningeal uptake.
Several approaches have been proposed to quantify regional tau pathology. While some have applied theory-driven regions of interest (ROIs) based on post-mortem findings (i.e., approximating the Braak staging scheme [11][12][13] for tau pathology), others have used approaches that are driven solely by the spatial patterns contained in tau PET images (i.e., data driven) [14][15][16][17]. As post-mortem and in vivo PET studies have shown that tau deposition patterns can deviate significantly from the Braak staging scheme [13,[18][19][20][21][22][23], datadriven approaches may provide a more accurate measure of tau burden. It is unclear, however, which of the two approaches (i.e., theory-or data-driven) is optimal for diagnostic purposes.
With an increasing number of available tau tracers, there will also be greater variability in quantitative outcome measures. This variability is due to the distinct properties of each tau tracer, and to different acquisition procedures across sites, analytical methods, and ROI selection. In contrast to amyloid PET, however, where outcome measures from different tracers or methods can be standardized to a common scale [24], tau PET findings cannot currently be directly comparable. A common scale for tau imaging would facilitate direct comparison of outcome measures and tracer characteristics, and help establish uniform cut-offs for early tau pathology and the range of tau positivity characteristic of AD.
Using [ 18 F]flortaucipir, [ 18 F]RO948, and [ 18 F]MK6240, the objectives of the present study were to compare the diagnostic performance (i.e., separating AD dementia from cognitively unimpaired (CU) individuals and non-AD disorders) of theory-and data-driven ROIs in order to examine whether a common target ROI (and cut-off to define tau PET positivity) can be used across tracers for differential diagnosis. In addition, at a broader level, comparison of findings across these three tracers can serve as a proof of concept with respect to the eventual standardization of tau PET imaging measures. . Groups were established without the use of biomarkers. We included only patients with Aβ-positive AD dementia in accordance with the National Institute on Aging-Alzheimer Association research framework [25]. Additional details on included cohorts and on the definition of study groups are included in Supplementary Tables 1-4. Informed consent was obtained from all participants, with studies approved by local institutional review boards.

Image acquisition and processing
Complete details on the acquisition and processing of tau PET data have been described elsewhere [9,11,[26][27][28] For all three tracers, images were first motion corrected, timeaveraged, and rigidly coregistered to their corresponding skull-stripped T1-weighted structural MRI scan. Standardized uptake value ratio (SUVR) images were created using the inferior cerebellar cortex as reference region for all tracers. FreeSurfer (v.6.0)-based parcellations of T1-weighted MRI scans were applied to the tau PET scans transformed to participants' native T1-space to extract mean regional SUVR values for each participant. For voxelwise analyses, SUVR PET images were spatially transformed into a common MNI152 space using the transformation derived from MRI normalization step.  [29]. In short, ETs are similar to the better-known Random Forest. As ETs fit each tree on all data without bagging, calculations were evaluated in 10-fold cross-validation loops yielding mean feature importance and standard deviations. Subjects of two subgroups (first AD dementia and CU, followed by AD dementia and non-AD) and their feature vectors (mean SUVR values in FreeSurfer ROIs) were used to train the ETs to predict group membership by decision tree majority vote and output the feature importance scores extracted from the ensemble. The highest ranked features from each resulting group were then extracted from the resulting dendrogram, first out of a single group (all features) and subsequently from the highest ranked groups in an iterative fashion. This allowed us to assess the relationship between accuracy and the number of group representatives added, providing a truly minimalist representation of group differentiating ROIs. The dendrograms resulting from this approach are provided for each tracer and contrast in Supplementary Fig. 1-6.

Theory-driven ROIs
In order to include brain areas affected by NFT pathology across the course of AD (i.e., from early to later affected areas), we created four FreeSurfer-based composite ROIs using an approach developed using [ 18 F]flortaucipir [11] and based on the Braak staging scheme for tau pathology [30]. These ROIs have been used previously [31,32] and include the entorhinal cortex (stage I/II), a temporal meta-ROI (entorhinal cortex, amygdala, inferior/middle temporal gyri, fusiform gyrus, and parahippocampal gyrus, approximating Braak I/IV) [33,34], and a neocortical meta-ROI (widespread neocortical areas, approximating Braak V/VI). In addition, we included an Early tau ROI comprising regions shown to accumulate tau early on in the course of AD (entorhinal cortex, i n f e r i o r t e m p o r a l c o r t e x , f u s i f o r m g y r u s , a n d parahippocampal gyrus) [35]. Despite substantial anatomical overlap with the temporal meta-ROI, the inclusion of this Early tau ROI was motivated by findings showing that tau PET signals in these different temporal regions follow different dynamics when it comes to tau accumulation, indicating that small differences in ROI composition may affect sensitivity [36].

Statistical analyses
All analyses were performed in R, v.4.0.2 (R Foundation for Statistical Computing, https:/d/www.R-project.org/), with significance set at P < 0.05, two-tailed. Demographics and tau PET SUVR values at the ROI level were compared across cohorts and diagnostic groups using analysis of variance and post hoc t-tests (continuous variables) or Fisher's exact tests (binary variables). Cut-offs for tau PET imaging ROIs were determined for each tracer using the Youden index (sensivity+specificity-1; AD dementia vs Aβnegative CU). The diagnostic performance of tau PET (AD dementia vs CU and non-AD) was assessed for each ROI by testing for significant differences between area under the receiver operating characteristic curve (AUC) values for ROIs using DeLong statistics [37].  Table 1 with average tau PET images across diagnostic groups and tracers shown in Fig. 1. SUVR values across tracers and diagnostic groups are shown for theory-driven ROIs in Fig. 2. For all three tracers, SUVR values across apriori ROIs were significantly higher in AD patients as compared to CU individuals and non-AD disorders (P < 0.001). For [ 18 F]flortaucipir, SUVR values in the non-AD group were significantly higher than those for CU individuals in the entorhinal cortex (P < 0.001), in the Early tau ROI (P < 0.05) and in the neocortical meta-ROI (P < 0.001). For [ 18 F]MK6240, SUVR values were significantly higher in CU individuals compared to non-AD in the Early tau and temporal meta-ROIs (Table 1).

Deriving data-driven ROIs
Extra-tree-based data-driven ROIs are shown by tau tracer in Fig. 3 Establishing cut-offs for data-and theory-driven ROIs Cut-offs for SUVR values within tau PET imaging ROIs (data-and theory-driven) were determined across tracers and are presented in Table 2. For the separation of AD dementia from non-AD disorders, the temporal meta-ROI showed the highest AUC for all three tracers; moreover, cut-off values appeared to converge around an SUVR of 1. 35  Diagnostic performance using theory-and datadriven ROIs AUC findings-along with sensitivity and specificity estimates-for theory-and data-driven ROIs are summarized in Table 2. When using theory-driven ROIs (i.e., entorhinal cortex, Early tau, temporal, and neocortical meta-ROIs) for  Table 6).

Discussion
In this multicentric study, we compared the cross-sectional diagnostic performance of [ 18 F]flortaucipir, [ 18 F]RO948, and tau PET for the separation of AD dementia from both CU individuals and non-AD disorders, using SUVR values drawn from both theory-driven (apriori) and data-driven ROIs. For these comparisons, no significant differences in AUC values were seen between the best performing theory-driven ROIs and those determined using the data-driven approach. Moreover, an SUVR value of approximately 1.35 appeared to be a common threshold across tracers for the temporal meta-ROI (corresponding to Braak I/IV). Ongoing work applying a functional connectivity-based approaches to define ROIs for tau PET suggests that datadriven approaches may better capture tau pathology crosssectionally and can provide patient-tailored ROIs that predict longitudinal tau accumulation with greater sensitivity than Braak-based stages [38]. Our findings suggest, however, that the temporal meta-ROI (approximating Braak stage I/IV) is suitable for achieving high diagnostic accuracy across tau PET tracers in differentiating AD dementia from CU individuals and non-AD dementia disorders. The ranking of the temporal meta-ROI as the best performing among apriori ROI is consistent with earlier studies using both [ 18 F]flortaucipir [31] and [ 18 F]RO948 [32]. Recent work using [ 18 F]flortaucipir has also highlighted the likely need for quantification as an adjunct to visual assessment if tau PET is to be used clinically [39]. In that study, physicians visually assessed [ 18 F]flortaucipir PET images as consistent or not consistent with AD. These ratings were then compared to tau immunohistochemistry and levels of AD neuropathologic change based on Aβ plaque burden [39]. Prespecified levels of sensitivity and specificity were not met by 2 of 5 readers, however, largely due to false-positive reads based on the misinterpretation of temporal lobe findings. Despite differences in study design, the present findings suggest that automated quantification of tau PET retention in the temporal meta-ROI could prove a suitable measure to support visual reads in clinical practice.
For the data-driven ROIs, overlap in the regions comprising the composite ROIs for both contrasts (i.e., AD dementia vs CU and non-AD) was high but imperfect across tracers. This finding reflects a combination of cohort effects (e.g., the amount and distribution of tau pathology across diagnostic groups within each cohort) and the use of different tracers.

W h i l e p o s t -m o r t e m s t u d i e s h a v e s h o w n t h a t [ 18 F]flortaucipir, [ 18 F]RO948, and [ 18 F]MK6240 all bind to
AD-type tau tangles [5,40,41], these tracers may differ in their sensitivity to other forms of tau pathology such as neuropil threads and dystrophic neurites. Future work using head-to-head study designs will be required to address this.
By applying a linear scaling operation to amyloid PET data, outcome data can be expressed in a common 100-point scale (unit termed "Centiloids") [24], with zero representing the average value seen in high-certainty Aβ-negative subjects and 100 the average in typical AD patient with mild-tomoderate dementia. Recent work applying this method to multicenter amyloid PET data [42] has shown the Centiloid approach to be feasible and robust; further, by incorporating comparisons with post-mortem measures of Aβ pathology, The regions that best separated AD dementia from CU individuals (parahippocampus and inferior temporal cortex) and those diagnosed with non-AD neurodegenerative disorders (entorhinal cortex, amygdala, parahippocampus, and inferior temporal cortex) using [ 18 F]flortaucipir PET. (B) The regions that best separated AD dementia from CU individuals (entorhinal cortex and amygdala) and those diagnosed with non-AD neurodegenerative disorders (entorhinal cortex, amygdala, parahippocampus, fusiform gyrus, and inferior temporal cortex) using [ 18 F]RO948. (C) The regions that best separated AD dementia from CU individuals (fusiform gyrus, inferior temporal cortex, middle temporal gyrus) and those diagnosed with non-AD neurodegenerative disorders (entorhinal cortex, amygdala, the inferior temporal cortex, the banks of the superior temporal sulcus, and the fusiform gyrus) using [ 18 F]MK6240 the authors reported a neuropathology-based Centiloid cut-off for amyloid PET positivity. The present study, while exploratory in nature, provides preliminary support for the development of a Centiloid-like methodology to facilitate comparisons across tau PET tracers. In the interim, our findings indicate that multicohort studies combining different tau tracers are possible if simply classifying particiapnts as tau positive (i.e., temporal meta-ROI > 1.35 SUVR) or negative. This study has limitations. First, the non-AD groups varied in size and composition. In the TRIAD cohort, for instance, there were no patients with DLB or svPPA. As both diagnoses have been associated with elevated tau PET signal [31,32,[43][44][45][46], this imbalance may explain the somewhat higher AUC values for [ 18  covering the spectrum around the proposed cut-off are required. The absence of Aβ-positivity in the non-AD group using [ 18 F]MK6240 may also explain the finding that SUVR values in the Early tau and temporal meta-ROIs were somewhat higher in the CU group, where 22% were amyloid PET positive [47]. Second, we here focused on differential diagnosis at the dementia stage. Subsequent studies using longitudinal data will be required to examine the role of tau PET in identifying CU and MCI individuals who subsequently progress to AD dementia. Though cohort differences precluded cross-cohort evaluations, the present study was not intended as a head-to-head comparison of the three tracers for diagnostic purposes. Third, these results may not generalize to other new tau PET tracers currently entering the field such as [ 18 F]PI-2620, [ 18 F]GTP1, or [ 18 F]JNJ-067 [4]. Lastly, though not the aim of the present study, these findings cannot be extrapolated with respect to what ROI(s)/cut-off(s) may prove best for longitudinal applications (e.g., predicting cognitive decline or tau accumulation). In addition to the large sample size, strengths of this study include the incorporation of both CU and non-AD subjects as diagnostic groups and the use of three different tau PET tracers.

Conclusion
A common-ROI encompassing parts of the temporal lobe (i.e., Braak I/IV) can be used for differential diagnosis of dementia patients with [ 1 8 F]

Declarations
Ethical approval Informed consent was obtained from all participants, with studies approved by local institutional review boards.
Conflict of interest Dr. Hansson reported receiving grants from Roche during the conduct of the study as well as grants from Roche, nonfinancial support from GE Healthcare, and grants from Biogen outside the submitted work. AL, TP, OS, PI, RS, NMC, ALB, HC, CHL, RLJ, GR, RO, and PRN report no conflicts of interest. Supplementary Information The online version contains supplementary material available at https://doi.org/10.1007/s00259-021-05401-4.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.