Development of software for measuring brain amyloid accumulation using 18F‐florbetapir PET and calculating global Centiloid scale and regional Z‐score values

Abstract Background and purpose Quantitative measures have been proposed to aid the visual interpretation of amyloid PET. Our objective was to develop and validate quantitative software that enables calculation of the Centiloid (CL) scale and Z‐score for amyloid PET with 18F‐florbetapir. Methods This software was developed as a toolbox in statistical parametric mapping 12 running on MATLAB Runtime. For each participant's amyloid PET, this software calculates the CL scale using the standard MRI‐guided pipeline proposed by the Global Alzheimer's Association Interactive Network (GAAIN) and generates a Z‐score map for comparison with a new amyloid‐negative database constructed from 20 healthy controls. In 23 cognitively impaired patients with suspected Alzheimer's disease, Z‐score values for a target cortical area from the new database were compared with those from the GAAIN database constructed from 13 healthy controls. The CL values obtained using low‐dose CT of PET/CT equipment were then compared with those obtained using MRI. Results The CL calculation was validated with the 18F‐florbetapir dataset in the GAAIN repository. Z‐score values obtained from the new database were significantly higher (mean ± standard deviation, 1.05 ± 0.77; p < .0001) than those obtained from the GAAIN database. The use of low‐dose CT provided CL scales that were highly correlated with those obtained with MRI (R 2 = .992) but showed a slight yet significant underestimation (−2.1 ± 4.2; p = .013). Conclusions Our quantification software provides the CL scale and Z‐score for measuring overall and local amyloid accumulation with the use of MRI or low‐dose CT.


INTRODUCTION
The clinical impact of amyloid PET on the diagnosis and patient management of Alzheimer's disease (AD) has been reported in many studies (Boccardi et al., 2016;Ceccaldi et al., 2018;Rabinovici et al., 2019;Zwan et al., 2017). In our recent multicenter study  using 18 F-florbetapir, amyloid PET results substantially changed the pre-scan AD/non-AD diagnosis and patient management plans in 39% and 42% of patients, respectively.
These changes were driven by visual interpretation of the positivity/negativity of amyloid PET. However, when amyloid accumulation is low, this dichotomized visual interpretation tends to vary among readers. For 18 F-florbetapir PET, the κ coefficient, which indicates inter-reader reliability, is high but ranges from just 0.69 to 0.74 (Camus et al., 2012;Nayate et al., 2015).
To aid the visual interpretation, quantitative measures of amyloid accumulation in the brain have been proposed. In particular, the Centiloid (CL) (Klunk et al., 2015) scale has become widely used in recent years as a harmonized value for standardizing each analytical method or PET ligand used. Several studies have reported CL thresholds for amyloid positivity. For example, comparative studies between antemortem PET and postmortem neuropathology reported that a CL less than 10 can exclude AD due to the absence of neuritic plaques (Amadoru et al., 2020) and that a cutoff of 12.2 CL detects moderateto-frequent neuritic plaques (La Joie et al., 2019). Positive visual interpretations are reported to be highly consistent for CL scales of 26 and above (Amadoru et al., 2020;, whereas CL scales from 12 to 30 are often equivocal findings on visual evaluation and are referred to as the gray zone (Milà-Alomà et al., 3021;Pemberton et al., 2022). As an adjunct to the visual detection of focal early amyloid accumulation in this gray zone, a comparison with a database of amyloid-negative controls, rather than with the whole brain CL scale alone, may be helpful (Lilja et al., 2016). We have already developed software, called Amyquant (Matsuda & Yamao, 2022) (Matsuda et al., 2021) and, in this study, we examine its application to 18 F-florbetapir.

Participants
The participants were 23 patients (14 women and 9 men; 75.4 ± 7.4 years old, range, 48-82 years) and 20 cognitively healthy adults (13 men and 7 women; 45.4 ± 3.9 years old, range, 35-50 years) enrolled in a previous multicenter study . The patients were recruited from an outpatient memory clinic of the National Center of Neurology and Psychiatry, Japan. They had a mini-mental state examination (MMSE) (Folstein et al., 1975)

Evaluation of Z-score values using a newly constructed database from healthy controls
The mean positive Z-score of the target cortical VOI obtained from AMYclz analysis based on the standard MRI pipeline in 23 patients was compared between that obtained using a newly constructed control database comprising 20 healthy controls and that obtained using a GAAIN database comprising 13 healthy controls. The mean positive Zscore values obtained using the new database and the GAAIN database were defined as Z new_db and Z GAAIN_db , respectively. In addition, SUVR values at the target cortical VOI were compared between the new database and the GAAIN database.

Evaluation of calculated CL scales and Z-score values using low-dose CT in patients
The CL scales calculated using AMYclz with the CT-guided pipeline were compared with those obtained using the MRI-guided pipeline in patients. The mean positive Z-score values of the target cortical VOI obtained from the new database were also compared between MRIguided and CT-guided pipelines in patients. The CL scales and mean positive Z-score values obtained using MRI-guided and CT-guided pipelines were defined as CL MRI , CL CT , Z MRI , and Z CT , respectively.

Statistical analysis
Concordances between Z new_db and Z GAAIN_db , between CL MRI and CL CT , and between Z MRI and Z CT were assessed using Pearson correlation estimates and Bland-Altman plots. In the Bland-Altman plot, we performed a Spearman correlation to test whether there were associations between the difference and the load. CL scales and Z-score values and their standard deviations were computed with mean absolute differences and limits of agreement. These statistical tests were performed using JMP ver. 16.2 (SAS Institute).

F I G U R E 1
Processing pipeline for the software for quantifying amyloid accumulation by 18 F-florbetapir PET one process entails calculation of the Centiloid (CL) scale from each participant's amyloid PET and MRI or CT. Another process involves the statistical comparison of each participant's amyloid PET with a database constructed from negative amyloid PET results obtained from healthy controls. These two processes are automatically executed sequentially as an SPM12 toolbox running in MATLAB Runtime on the Windows operating system.
In addition, to investigate regional differences in the CT-guided and MRI-guided standardized amyloid PET images, a paired t-test was applied to these images on a voxel basis after they were smoothed with an 8-mm FWHM Gaussian kernel using SPM12. Results were considered significant with an extent threshold of 300 voxels corrected for multiple comparisons (family-wise error [Flandin & Friston, 2019], p < .05).

Validation of the present software for CL calculation
Validation of the present processing pipeline using 46 pairs of 18 Fflorbetapir PET and corresponding 3DT1WI datasets in the GAAIN repository indicated an excellent correlation with published data. The slope of the linear correlation was 1.0, with an intercept of 0.293, and the R 2 was .997, which were within the validation criteria defined by Klunk et al. (2015). These criteria state that the slope should be between 0.98 and 1.02 and the intercept between −2 and +2 CL for a linear regression equation and that the R 2 correlation coefficient should exceed 0.98. This validation of the present pipeline allowed use of the previously published equation (Navitsky et al., 2018) for the direct conversion of the 18 F-florbetapir SUVR to CL.

Evaluation of Z-score mapping using a newly constructed database of cognitively healthy controls
Pearson correlation analysis revealed a highly significant correlation of R 2 = .997 between Z new_db and Z GAAIN_db (p < .0001, Figure 2a). The linear regression equation was Z new_db = 1.393 × Z GAAIN_db + 0.04.
A Bland-Altman plot showed that Z new_db was significantly higher F I G U R E 2 Comparison of the Centiloid (CL) scale and Z-score values between different databases and between MRI-guided and CT-guided pipelines. Pearson correlation analysis (a, c, e) showed highly significant correlations (p < .0001) of R 2 = .997 between Z new_db and Z GAAIN_db , of R 2 = .992 between CL CT and CL MRI , and of R 2 = .994 between Z CT and Z MRI . A Bland-Altman plot (b, d, f) showed that Z new_db was significantly higher than Z GAAIN_db (mean ± standard deviation, 1.05 ± 0.77; p < .0001), that CL CT was slightly but significantly underestimated (−2.1 ± 4.2; p = .013) compared with CL MRI , and that the positive mean Z-score for the target cortical volume of interest (VOI) was not significantly different between Z CT and Z MRI (p = .253). Spearman correlation analysis revealed a significant association between the difference in Z new_db versus Z GAAIN_db and Z load (ρ = .955, p < .0001), no significant association between the difference in CL CT versus CL MRI and CL load (ρ = −.083, p = .701), and a significant association between the difference in Z CT versus Z MRI and Z load (ρ = −.460, p = .027).
than Z GAAIN_db (mean ± standard deviation, 1.05 ± 0.77; p < .0001, Figure 2b). The 95% limits of agreement ranged from 0.7 to 1.4. Spearman correlation analysis revealed a significant association between the difference in Z new_db versus Z GAAIN_db and Z load (ρ = .955, p < .0001). The mean and standard deviation images for the new database and GAAIN database are demonstrated in Figure 3. SUVR values in the target cortical VOI were significantly lower in the new database (0.986 ± 0.052) than in the GAAIN database (1.048 ± 0.064) (p < .005). Results of AMYclz analysis using the new database and GAAIN database are presented for a representative case (Figure 4).

F I G U R E 3
Standardized uptake value ratio (SUVR) images for the mean and standard deviation of the present new database and the Global Alzheimer's Association Interactive Network (GAAIN) database. In the target cortical volume of interest (VOI), the SUVR values (0.986 ± 0.052) for the present new database (new_db, 13 men and 7 women; 45.4 ± 3.9 years old) are significantly (p < .005) lower than those (1.048 ± 0.064) for the GAAIN database (GAAIN_db, 7 women and 6 men; 27.0 ± 4.3 years old).

Evaluation of the calculated CL scales and Z-score using low-dose CT in patients
MRI-guided and CT-guided anatomically standardized PET images ( Figure 5) are presented, along with results of AMYclz analysis, for a representative case. Pearson correlation analysis showed a highly significant correlation of R 2 = .992 between CL CT and CL MRI (p < .0001, Figure 2c). The linear regression equation was CL CT = 0.989 × CL MRI − 1.68. A Bland-Altman plot showed that CL CT was slightly but significantly underestimated (mean ± standard deviation, −2.1 ± 4.2; p = .013) compared with CL MRI (Figure 2d). The 95% limits of agreement ranged from −3.9 to −0.3. Spearman correlation analysis did not reveal a significant association between the difference in CL CT versus CL MRI and CL load (ρ = −.083, p = .701).
Pearson correlation analysis revealed a highly significant correlation of R 2 = .994 between Z CT and Z MRI (p < .0001, Figure 2e). A Bland-Altman plot showed that the positive mean Z-score for the target cortical VOI was not significantly different between Z CT and Z MRI (p = .253, Figure 2f). Spearman correlation analysis identified a significant association between the difference in Z CT versus Z MRI and Z load (ρ = −.460, p = .027).
Paired t-tests performed using SPM12 ( Figure 6 and Table 1) found that the brainstem exhibited the biggest differences in uptake between the CT-guided and MRI-guided standardized PET images. Significantly lower uptake of CT-guided standardized PET images versus MRIguided PET images was observed in the frontal cortex of the target cortical VOI.

DISCUSSION
In this study, we developed and validated new quantification software for amyloid PET with 18 F-florbetapir. This software can not only calculate the CL scale at the target cortical VOI of the GAAIN database using MRI-guided or CT-guided pipelines but also provide a Z-score map for a

F I G U R E 4 Comparison of analytical results by AMYclz between the new database and Global Alzheimer's Association Interactive Network
(GAAIN) database. A Z-score map is displayed by overlay on tomographic sections with a contour of the target cortical volume of interest (VOI) and with a surface rendering of the standardized brain MRI. A regionally higher Z-score was obtained from the present new database (left) than from the GAAIN database (right). Previous studies have examined the use of low-dose CT of PET/CT equipment as a substitute for MRI in the calculation of the CL scale (Kim et al., 2022;Matsuda et al., 2021;Presotto et al., 2018). Our study using 18 F-flutemetamol revealed that low-dose CT provided a CL scale comparable to that of MRI. However, the CL scale obtained with low-dose CT was on average 1.7 points lower than that obtained with MRI. In the present study, conducted using 18 F-florbetapir, low-dose CT also showed a CL scale value that was 2.1 points lower on average than that obtained with MRI. Nevertheless, the CL scale obtained with low-dose CT was highly correlated with the gold standard CL scale obtained with MRI within the validation criteria proposed by Klunk et al. (2015). The reason why the CT CL was slightly lower than the CL MRI is that the CT-guided anatomically standardized PET images showed a lower accumulation in the target cortical VOI than the MRIguided standardized PET images. This may be due to the slightly lower accuracy of anatomic standardization with low-dose CT compared with MRI. On the other hand, the mean positive Z-scores for the target cortical VOI compared with the new database tended to be lower for Z CT than Z MRI in the high range but; overall, there was no significant difference between Z CT and Z MRI . In addition, in this regard, low-dose CT F I G U R E 5 Comparison of analysis results from AMYclz CT-guided and MRI-guided pipelines (top). Anatomically standardized MRI, low-dose CT, and amyloid PET images are shown; MRI-guided and CT-guided anatomical standardization resulted in nearly identical PET images. (Bottom) Results of AMYclz analysis of CT-guided and MRI-guided pipelines are shown; the CT-guided pipeline (left) yielded a slightly lower CL65.1 than the CL67.1 obtained for the MRI-guided pipeline (right), but the Z-score maps were identical for these two pipelines.

Cluster size T-Value
of PET/CT equipment may be a potential substitute for MRI in amyloid PET quantification if MRI is not obtained in the same period as PET.
In contrast, several studies (Bourgeat et al., 2018;Edison et al., 2013;Fujishima & Matsuda, 2022;Imabayashi et al., 2022;Saint-Aubert et al., 2014;Tsubaki et al., 2020), including our own research (Fujishima & Matsuda, 2022), have reported anatomic standardization using the amyloid PET template alone without MRI. Out of these studies, the uses of an adaptive template generated from a linear combination of an amyloid-negative and amyloid-positive template with a weight-optimized algorithm have obtained comparable CL scales to those obtained with the standard MRI-guided method (Bourgeat et al., 2018;Fujishima & Matsuda, 2022;Imabayashi et al., 2022). However, our correlation analysis (Fujishima & Matsuda, 2022)  accumulation. In such cases, anatomic standardization using coregistered structural images may be more accurate than the PET-alone method.
The limitation of this study is the small number of cases for the comparison of CL MRI and CL CT . Further studies with more cases are needed to assess the lower limit of the dose at which the algorithm would function properly with a reduced current and whether the results would be improved by the use of diagnostic-quality CT which increases head radiation dose to about 3 mSv. The software provides SUVR and CL scales for the entire target cortical VOI, but it may be necessary to calculate these values for segmented regions.
In conclusion, our newly developed amyloid PET quantification software for 18 F-florbetapir, named AMYclz, provides the CL scale and Z-score to quantify the overall and local amyloid accumulation for a patient's PET compared with an amyloid-negative database comprising 20 healthy controls. This software supports the use of low-dose CT in PET/CT equipment, in addition to the standard use of MRI in the anatomical standardization of amyloid PET.

F I G U R E 6
Direct comparison of anatomically standardized amyloid PET images using low-dose CT and MRI statistical parametric mapping (SPM) analysis showed significantly (family-wise error, p < .05) higher and lower uptake of CT-guided standardized PET images than MRI-guided standardized PET images, presented in a warm color scale and a cool color scale, respectively. Volume of interest (VOI) templates are shown as a solid black area for the whole cerebellum as a reference area and as a solid white area for the target cortical area. The largest differences in accumulation are visible in the brain stem. Lower uptake of CT-guided standardized PET images was observed in the frontal cortex of the target cortical VOI.

ACKNOWLEDGMENTS
We thank all clinicians and imaging technicians who contributed to this study.

CONFLICT OF INTEREST STATEMENT
Tsutomu Soma is an employee of PDRadiopharma Inc. Kyoji Okita has received a research grant from PDRadiopharma Inc. The other authors declare no conflict of interests.

DATA AVAILABILITY STATEMENT
Research data are not shared.