Combination of automated brain volumetry on MRI and quantitative tau deposition on THK-5351 PET to support diagnosis of Alzheimer’s disease

Imaging biomarkers support the diagnosis of Alzheimer’s disease (AD). We aimed to determine whether combining automated brain volumetry on MRI and quantitative measurement of tau deposition on [18F] THK-5351 PET can aid discrimination of AD spectrum. From a prospective database in an IRB-approved multicenter study (NCT02656498), 113 subjects (32 healthy control, 55 mild cognitive impairment, and 26 Alzheimer disease) with baseline structural MRI and [18F] THK-5351 PET were included. Cortical volumes were quantified from FDA-approved software for automated volumetric MRI analysis (NeuroQuant). Standardized uptake value ratio (SUVR) was calculated from tau PET images for 6 composite FreeSurfer-derived regions-of-interests approximating in vivo Braak stage (Braak ROIs). On volumetric MRI analysis, stepwise logistic regression analyses identified the cingulate isthmus and inferior parietal lobule as significant regions in discriminating AD from HC and MCI. The combined model incorporating automated volumes of selected brain regions on MRI (cingulate isthmus, inferior parietal lobule, hippocampus) and SUVRs of Braak ROIs on [18F] THK-5351 PET showed higher performance than SUVRs of Braak ROIs on [18F] THK-5351 PET in discriminating AD from HC (0.98 vs 0.88, P = 0.033) but not in discriminating AD from MCI (0.85 vs 0.79, P = 0.178). The combined model showed comparable performance to automated volumes of selected brain regions on MRI in discriminating AD from HC (0.98 vs 0.94, P = 0.094) and MCI (0.85 vs 0.78; P = 0.065).


Scientific Reports
| (2021) 11:10343 | https://doi.org/10.1038/s41598-021-89797-x www.nature.com/scientificreports/ underlying pathophysiologic culprits of AD, namely senile plaques and neurofibrillary tangles, but their clinical use may be affected by its availability, cost and consideration of ionizing radiation exposure. MRI is used for standard practice to support the diagnosis of AD and to exclude other causes of cognitive impairment such as stroke, vascular dementia, normal pressure hydrocephalus, or inflammatory or neoplastic conditions. Atrophy of the medial temporal lobe, particularly the hippocampus and entorhinal cortex, was shown to be important in patients on the AD spectrum and these regions were identified as predictors of time to progression from mild cognitive impairment (MCI) to AD [2][3][4][5][6] . Qualitative visual assessment of atrophy on MRI suffers from poor interobserver agreement, which is a major limitation for implementing it in a clinical setting 7,8 . This limitation can be addressed by quantitative volumetry analysis, and automated software programs for quantitative volumetric analysis have been developed 9,10 . Extensive validation of these software packages has been done for clinical use with comparison to manually segmented volumetric measurements [11][12][13][14] , and commercially available automated brain volumetric tool showed high correlation in brain volume measurement with tools extensively used in research settings with moderate to high sensitivity (63.3-83%) and high specificity (93-100%) in differentiating AD from healthy controls (HC) [15][16][17][18] .
[18F] THK-5351 is one of the first generation of tau PET tracers 19 and a close correspondence between the Braak staging of tau pathology and retention of the first generation tau PET tracers has been demonstrated in AD 20,21 . Braak stages of tau pathology, derived from cross-sectional data, proposed how AD-related tau pathology begin in medial temporal structures extending to limbic areas, posterior cingulate cortex and then widely to isocortical brain areas [22][23][24] . While tracer specificity related to off-target binding of [18F] THK5351 to monoamine oxidase B (MAO-B) remains as a possible limitation, previous studies reported that binding of [18F] THK5351 in the hippocampus was not influenced by off-target binding in the choroid plexus unlike [18F] AV-1451 19,[25][26][27][28][29] . Moreover, longitudinal analysis of tau PET demonstrated that measurement of changes in the tau PET SUVR can be used as an efficient outcome measure in disease modifying clinical trials 30,31 .
While tau PET reflects the underlying pathophysiologic hallmark of AD, MRI demonstrates regional atrophy with automated brain volumetry extensively validated for clinical use in AD. To our knowledge, no direct comparison has been performed between commercially available automated brain volumetry on MRI and [18F] THK-5351 PET in discrimination of AD spectrum. The purpose of this study was to compare commercially available automated brain volumetry on MRI and [18F] THK-5351 PET, and determine if the combination of these tools can aid discrimination of AD spectrum in a clinically feasible setting.

Methods
All the experimental protocols were approved by institutional review boards of Asan Medical Center and Samsung Medical Center, Seoul, South Korea. All relevant study protocols for involving humans were in accordance with guidelines of institutional ethics committees.
Participants. From a prospective cohort of a multicenter clinical trial (NCT02656498), 113 participants who had both baseline structural MRI and [18F] THK-5351 PET were included. All participants or their appropriate representatives provided informed consent, and all participants were examined under the protocols approved by the institutional review board of the two tertiary medical centers (Asan Medical Center and Samsung Medical Center). A flow diagram of participant inclusion is shown in Fig. 1 36,37 .
Prior to the PET scan, a head holder was applied to minimize head motion and brain CT images were obtained for attenuation correction. Using the ordered-subsets expectation maximization algorithm (iteration = 4, subset = 16), 3D PET images were reconstructed with a voxel size of 2.0 × 2.0 × 3.27 mm 3 . To increase data uniformity between different PET scanners, a 3D Hoffman phantom-based PET harmonization method was applied based on the results of correction factors obtained from phantom studies using 3D Hoffman brain phantom 38,39 .
Quantitative image analysis. MRI volumetry. Automated MRI-volumetry analyses using NeuroQuant software package (CorTechs Labs, La Jolla, CA, USA) were performed via the standard processing pipeline. The details of this procedure are previously described elsewhere 11,40 . DICOM files were uploaded to the servers for processing. Briefly, the protocol entails a quality check, adjustment for gradient non-linearity/B1 field inhomogeneity, and skull stripping followed by a discrete cosine transformation and registration onto a dynamic probabilistic atlas (https:// www. corte chsla bs. com/ white papers/). An anatomic label is designated to each voxel based on approximations from the dynamic probabilistic atlas that is structure-wise similar to the FreeSurfer but uses an independent code-based method for intensity normalization and gradient distortion correction to accommodate scanner-specific acquisition-level differences and to better represent the aged population. The MRI analysis provides a report that includes raw and corrected volumes (% intracranial volume) for 66 brain regions. All reports were reviewed to ensure adequate quality.
[18F] THK-5351 PET images were segmented by using the default automated gyral-based parcellation method of FreeSurfer (version 5.3.0; http:// surfer. nmr. mgh. harva rd. edu) 13,24 . The SUVR of the cerebral cortex was calculated using the SUV of the cerebellar cortex as a reference region. Bilateral Braak stage region of interest (ROI) were created by combining FreeSurfer ROIs into non-overlapping Braak regions, using the Desikan-Killiany cortical atlas 41 . Further analyses from the automatic FreeSurfer segmentation were utilized for the left and right hippocampus. The values for the left and right hemispheres were then averaged. The ROIs of the in vivo Braak staging were expressed as Braak I/II, Braak III/IV and Braak V/VI, reflecting the propagation of tau pathology, which begins in the medial temporal structures, extends to limbic areas and then widely spreads into isocortical brain areas [22][23][24]

Statistical analysis. Comparison of demographics and quantitative values.
The clinical characteristics of the participants in each group (HC, MCI, and AD) were compared. Chi-square tests were used for categorical variables and analysis of variance (ANOVA) was used for continuous variables with post hoc pairwise comparisons where relevant. The cortical volumes and SUVRs of Braak ROIs were subsequently compared between groups using ANOVA, and baseline age was used as a covariate. Post-hoc analysis between groups was also conducted. In order to account for multiple comparisons across the multiple cortical regions, false discovery rate (FDR)-corrected P-value < 0.05 was used. P values for SUVRs were Bonferroni-corrected. Normality of all ROIs and Braak staging was verified using Kolmogorov-Smirnova and Shapiro-Wilk tests.
Construction of a diagnostic model. In order to identify significant brain regions in differentiating AD from HC and MCI, univariate and multivariate logistic regression analysis were performed using a general linear model with baseline age as a covariate. Stepwise logistic regression analyses were performed, and regions identified as significant regions upon univariate analysis underwent multivariate fitting to determine the best combination of regions for discriminating AD from HC and MCI. Variance inflation factor was considered to check for multicollinearity of the variables included in the regression equations and variables with variance inflation factor greater than 5 were excluded. The volume of the hippocampus was included as selected brain regions as the hippocampus is a well-established brain region that is known to be affected in AD and its atrophy is one of the most commonly used imaging criteria used for diagnosing AD 2,6,43 . The combined model was constructed using volumes of selected brain regions and SUVRs of Braak ROIs.
Performance measurement. Performance of discriminating AD from HC and MCI was quantified by area under the receiver operating characteristics (AUROC). The performances of volumes of selected brain regions, SUVRs of Braak ROIs and a combined model were calculated. The performance of the hippocampus alone was also evaluated.
Additional analysis. In order to provide the same ROI categorization for automated volumes on MRI and SUVRs on [18F] THK-5351 PET, automated volumes on MRI were grouped into Braak I/II, Braak III/IV and Braak V/VI as previously described, and univariate and multivariate logistic regression analysis on SUVRs from [18F] THK-5351 PET was performed as previously described. Performance of automated volumes on MRI grouped into Braak ROIs and SUVR of selected brain region was also evaluated.
Statistical analyses were performed using R statistical software (version 3.3.3, R Core Team, Vienna, Austria). Figure 2 shows the overall process for image and statistical analysis.

Results
Participants. The demographics of the study participants are provided in Table 1. Of note, the mean age in the AD group was lower than in the HC and MCI groups (P = 0.001). The sex ratio, education level, and APOE allele variations were comparable across the three groups. In particular, there was no difference in the proportions of participants who are either homozygotes or heterozygotes for APOE allele variations between the AD and MCI groups [31% (8/26) vs 40% (22/55); P = 0.437]. As expected, the K-MMSE score was lower in the AD group than in the other groups (P < 0.001). The mean interval between [18F] THK-5351 PET and MRI was 53.6 days (range, 0-202 days; standard deviation, 47 days) with no significant difference among the three groups.
SUVRs of Braak ROIs were also compared and post-hoc analysis was performed between groups (Supplementary Table 1

Discussion
In this study, performances of automated brain volumetry on MRI and quantitative measurement of tau deposition on [18F] THK-5351 PET were evaluated in discriminating AD on the AD spectrum. The combined model incorporating automated volumes of selected brain regions on MRI (cingulate isthmus, inferior parietal lobule, hippocampus) and SUVRs of Braak ROIs on [18F] THK-5351 PET showed higher performance than SUVRs   Table 3. Comparison of combined and individual performance using volumes of selected brain regions, SUVRs of Braak ROIs and volume of the hippocampus in discriminating AD from HC and MCI. Numbers in parentheses are 95% confidence intervals. AUROC = area under the receiver operating characteristics curve. P < 0.05 values are indicated in bold *P-value refers to the significance between the differences of the AUROCs between the combined model and other model.   While tau PET reflects the underlying pathophysiologic mechanism of AD, MRI demonstrates regional atrophy for which automatic volumetric assessment tool is available. In our study, commercially available, FDAapproved automated brain volumetry tool was evaluated that is being increasingly used in clinical settings as it is easily accessible providing fully automated service in a clinically acceptable time-period and feasible setting. Diagnostic efficacy of the automated volumetric assessment tool has previously been demonstrated in patients on the AD spectrum with high sensitivity (63.3-83%) and high specificity (93-100%) in differentiating AD from HC 5,6 , which is comparable to the performance demonstrated in our study (AUROC, 0.94; sensitivity, 88.5%; specificity, 90.6%). Previous studies used volume of the medial temporal lobe in the discrimination of AD while our study identified volumes of the cingulate isthmus and inferior parietal lobule by performing stepwise logistic regression analyses. The value of automated volumetric measurement of both the cingulate isthmus and inferior parietal lobule has been previously reported 6,16 , and the combination of automated volumetry of the entorhinal cortex and inferior parietal lobule was the best predictor of time to progress from MCI to AD 6 . Preferential atrophy of specific laminae in the inferior parietal lobule has been demonstrated in the early stages of AD on pathologic studies 44,45 . In addition, projections from the inferior parietal lobule include subfields within the medial temporal lobe reflecting spread of AD pathology from the temporal lobe to an interconnected region in the parietal lobe 46 .
While automated brain volumetry on MRI may be easily accessible in daily clinical practice, most studies have been performed in a research setting for tau PET. Currently there is no consensus for quantifying tau deposition on tau PET, and multiple tau-PET quantification methods (in vivo Braak staging, regional uptake in Braak composite regions, several whole-brain measures of tracer uptake, regional uptake in AD-vulnerable voxels and uptake in a priori defined regions) showed that all methods were related to amyloid and global cognition but regional measures covering AD-vulnerable regions increased sensitivity to early tau PET signal, atrophy and memory decline 47 . We based our analysis of [18F] THK-5351 PET according to the methods described by Scholl et al. 24   www.nature.com/scientificreports/ to MAO-B with greatest reductions seen in SUVs in the basal ganglia and thalamus when MAO-B inhibitors were used 42 . While the heterogeneous MAO-B availability across the cortex may limit the interpretation of [18F] THK-5351, studies have demonstrated that no statistically significant reductions were seen using standard reference region-based approach 29 . In addition, significant tau tracer retention in the temporal lobe as well as extra-temporal regions has been reported even in cognitively normal older population supporting the primary age-dependent tauopathy despite some controversy regarding this entity 24,48,49 . There were several limitations in our study. First, there were relatively small number of patients with AD (23%, 26/113) and overrepresentation of MCI (49%, 55/113). Second, age correction was not performed in the quantitative analysis of Braak ROIs from [18F] THK-5351 PET. It has been shown that tau PET SUVR demonstrated modest association with age throughout most regions of the brain in HC 48 but currently there is no consensus or established methods and reference standards for age correction in the quantitative analysis of Braak ROIs. Most relevant studies relate to [18F] AV-1451 and further work is required pertaining this issue in relation to [18F] THK-5351. Third, the mean age in the AD group was significantly lower than in the HC and MCI group suggesting the inclusion of early onset AD patients in the study. However, despite alleged difference in the pathogenesis and clinical features between early onset AD and late onset AD, they are known to be markedly similar in terms of their biological profiles including abnormalities in amyloid and tau biomarkers 31 . Despite inclusion of early onset AD patients, there was no difference in the proportion of patients who were homozygotes or heterozygotes for APOε4 allele, which may be due to small sample size in the AD group and selection bias. Fourth, previous study showed that there were significant differences in automated volumetric measurement across imaging sites which may be due to variability in scanner characteristics 50 and reproducibility for automated volumetric analysis was not tested.
In conclusion, this study demonstrated the robust performance of automated brain volumetry on MRI and quantitative measurement of tau deposition on [18F] THK-5351 PET in discrimination of AD spectrum with added value of automated brain volumetry in discriminating AD from HC. The cingulate isthmus and inferior parietal lobule were identified as significant brain regions in discriminating AD on the AD spectrum, and may be used as adjunct to the traditionally used hippocampal volume.