Development and Clinical Validation of CT-Based Regional Centiloid Method for Amyloid PET

Background: We developed and validated CT-based regional Centiloid. A CT-based regional Centiloid was developed and validated in the present study. Methods: For development ofMRI-based or CT-based regional CLs,the cohort consist of 63 subjects (20 young controls (YC) and18 old controls (OC), and 25 Alzheimer’s disease dementia (ADD)).We used direct comparison of FMM-FBB CL (dcCL) method using MRI and CT images to dene a common target region and six regional VOIs including the frontal, temporal, parietal, posterior cingulate, occipital and striatal regions. Global and regional dcCL scales were compared between MRI-based and CT-based methods. For clinical validation, cohortconsisted of 2,245subjects (627 in CN, 933 in MCI, and 685 in ADD). Results: Both MRI-based and CT-based dcCL scales showed that FMM and FBB were highly correlated with each other, globally and regionally (R 2 = 0.96~0.99). Both FMM and FBB showed that CT-based regional dcCL scales were highly correlated with MRI-based regional dcCL scales (R 2 = 0.97~0.99). Absolute differences in regional CL scales between CT-based and MRI-based methods seemed to be relatively insignicant (p>0.05). In our clinical validation study, the G(-)R(+) and G(+)Str(+) groups predict worse neuropsychological performance than the G(-)R(-) and the G(+)Str(-) groups (p< 0.05) respectively. Conclusions: Our ndings suggested that it is feasible to convert FMM or FBB dcSUVR values into the dcCL scales regionally without additional MRI scans, which might in turn become a more easily accessible method for researchers and be applicable to a variety of different conditions. and striatal (-) Aβ dcCL scales; G(+)Str(+): global (+) and striatal (+) Aβ dcCL scales; PC: posterior cingulate; aMCI: amnestic mild cognitive impairment; CN: cognitive normal; dcCL; CDR-SOB: clinical dementia rating scale sum of boxes; SVLT delayed recall: seoul verbal learning test delayed recall;


Background
The standard Centiloid (CL) method was recently proposed to harmonize and quantify 18 F-labeled amyloid beta (Aβ) PET ligands using 11 C-labeled Pittsburgh compound B ( 11 C-PiB) images as a reference [1][2][3][4][5]. The equations derived for conversion of SUVR into CL scales in previous studies used 11 C-PiB PET and 18 F-labeled amyloid PET images and applied the equations to 18 F-labeled Aβ ligands to convert standard CL scales using the 11 C-PiB ligand [1][2][3][4][5]. However, 11 C-PiB PET ligands are not available in most medical centers due to the limitations described above. Therefore, in our previous study, a direct comparison CL (dcCL) method that harmonizes 18 F-orbetaben (FBB) and 18 F-utemetamol (FMM) PET ligands without 11 C-PiB images was developed [6].
Because PET has low spatial resolution, MRI can be used as anatomical reference to quantify PET uptakes. Therefore, both the standard CL process and the dcCL method require PET and MR images to measure CL scales. However, in practical assessments, MRI scans can be a risk for certain patients who have devices such as cardiac pacemakers or implantable cardioverter de brillators (ICDs) [7] and receiving both PET and MRI scans may be a nancial burden. Because PET and CT images can be acquired through only one scan in the PET-CT scanner, the CT images can be used as anatomical reference images to PET images instead of MRI.
Both the standard CL and the dcCL methods generate global CL scales but not regional brain Aβ uptake information. However, in our previous study, regionally increased Aβ uptakes were shown associated with cognitive impairments [8]. In particular, patients with striatal involvement of Aβ showed a worse prognosis [9]. Therefore, harmonizing Aβ uptakes between different Aβ ligands globally and regionally for earlier detection and better prediction of prognosis is needed.
In the present study, CT-based regional dcCL scales were developed using a head-to-head comparison cohort of patients who underwent FMM and FBB. Because our MRI-based global dcCL method previously showed the two ligands were mutually highly correlated, we hypothesized that MRI or CT-based regional dcCL scales of FMM would correlate with FBB. In addition, hypothetically, our CT-based regional dcCL scales are comparable to MRI-based regional dcCL scales regarding reliability and precision. Finally, to validate the clinical e cacy of the newly developed CT-based regional dcCL scales, the effects of regionally increased Aβ uptakes on cognitive impairments, including striatal involvement of Aβ, were explored.

Participants
To develop MRI-or CT-based global and regional dcCLs, the study cohort included 63 subjects; 20 young controls (YCs), 18 old controls (OCs), and 25 individuals with Alzheimer's disease dementia (ADD). The subjects underwent paired FMM and FBB PET-CT and three-dimensional (3D) T1 MRI. Healthy YCs were under 40 years of age with normal cognitive function and no history of neurological or psychiatric disorders. OCs were over 65 years of age with normal cognitive function determined using neuropsychological tests and no history of neurological or psychiatric disorders. Participants diagnosed with MCI had to meet Petersen's criteria [10] and show objective memory impairment one standard deviation (SD) below the norm in at least one memory test. ADD was diagnosed based on the National Institute on Aging-Alzheimer's Association (NIA-AA) research criteria for probable AD [11].
All participants underwent clinical interviews, neurological and neuropsychological examinations, and laboratory tests including complete blood count, blood chemistry, thyroid function tests, syphilis serology, and vitamin B12/folate levels. The absence of structural lesions including cerebral infarctions, brain tumors, vascular malformations, and hippocampal sclerosis was con rmed based on brain MRI.
The Institutional Review Board of Samsung Medical Center approved the study protocol and all methods were performed according to the approved guidelines. Written consent was obtained from each participant.

MRI data acquisition
Standardized 3D T1 turbo eld echo images were acquired from all participants at Samsung Medical Center using the same 3.0 T MRI scanner (Philips Achieva; Philips Healthcare, Andover, MA, USA). The detailed parameters are described in Additional le 1, Supplementary Methods 1.
Aβ PET-CT data acquisition Participants underwent FMM and FBB PET at Samsung Medical Center using a Discovery STe PET/CT scanner (GE Medical Systems, Milwaukee, WI, USA) in 3D scanning mode that examined 47 slices 3.3mm in thickness spanning the entire brain [12,13]. Paired FMM and FBB PET images were acquired on two separate days; the mean interval time (4.0 ± 2.5 months across all groups) was not different among the three groups (p = 0.89). Among the 63 head-to-head dataset, FMM PET was performed rst in half of the patients (total 36: 19 ADD, 6 OCs, and 11 YCs) and FBB PET rst in the other half of the participants (total 27: 6 ADD, 12 OCs, and 9 YCs). According to the protocols for the ligands proposed by the manufacturers, a 20-min emission PET scan with dynamic mode (consisting of 4 × 5 min frames) was CT images were acquired using a 16-slice helical CT (140 KeV, 80 mA; 3.75-mm section width) for attenuation correction and reconstructed in a 512 x 512 matrix with a voxel size of 0.5 mm x 0.5 mm x 3.27 mm.
Regional visual assessment and regional dcCL scales Two experienced neurologists visually quanti ed FMM and FBB images [14] in our head-to-head dataset.
Each doctor scored the frontal, parietal, posterior cingulate/precuneus, lateral temporal, and striatum as positive or negative and recorded the overall amyloid status. Inter-rater agreement was excellent for FMM (Fleiss k = 0.86-0.97) and FBB (Fleiss k = 0.9-1.0) for ve regions. After individual ratings were performed, the nal visual positivity was determined based on the majority of agreement regarding visual reading results.

Development of MRI-based global and regional CTX VOI
The method overview is shown in Fig. 1. To develop MRI-based dcCLs, we followed CL process for preprocessing (i.e., creating SUVR parametric PET images on MNI space) to MRI, FMM, and FBB PET images as described in Klunk et al. The processing details are provided in the original CL manuscript [3].
Brie y, individual MR images were co-registered onto the MNI-152 template and then individual PET images were co-registered onto the corresponding MRI images. The PET and MRI images were spatially normalized ( Fig. 1b) using transformation parameters of SPM8 uni ed segmentation method of T1weighted MRIs. The whole cerebellum (WC) mask was used as the reference region from the GAAIN website and SUVR parametric PET images of FMM and FBB using the WC mask created and used for global and regional dcCLs. To de ne the common cortical target region with amyloid accumulation distributed for FMM and FBB PET, 25 Aβ PET-positive (+) ADD patients and 18 Aβ PET-negative (-) OCs were included for all PET ligands in head-to-head cohort [6]. The method in the original publication was used to create a common cortical target VOI (FMM-FBB global CTX VOI) for both FMM and FBB PET and the details are described in the original publication ( Fig. 1d1) [6]. Individual dcSUVR values in the FMM-FBB global CTX VOI were calculated in all PET images. MRI-based regional VOIs were de ned by overlapping MRI-based FMM-FBB global CTX VOI and AAL atlas (Fig. 1e1) [15]. The sub-regions of AAL in the MRI-based global CTX VOI were merged into six regions (frontal, PC, parietal, striatum, occipital, and temporal). Regional dcSUVR values were calculated using the six regional VOIs.
Development of CT-based global and regional CTX VOI For constructing the brain CT template, 139 CT scans were collected from normal controls (NCs) in another dataset who underwent FBB-PET CT. In the PET-CT scanner, the CT image was low-dose CT to reduce patient exposure to radiation. Brain CT template was created using corrected Houns eld units (HU) of brain tissues in the CT images. The details of the HU correction approach are described in the original methodology paper [16]. Brie y, as shown in Fig. 1a, total CT images were reoriented. Intensities of the images were scaled to boost HU of brain tissues. The HU-corrected CT images were co-registered onto corresponding T1 MR images. Individual T1 MR images were spatially normalized on MNI space and spatial normalization parameters of T1 MR images were applied to corresponding HU-corrected CT images. The normalized CT images were ipped to create a symmetric template and the mean image was created using the normalized CT images. Gaussian smoothing at 8 mm was applied to the template to remove template noise.
HU correction processing was performed to individual CT images. FMM and FBB PET images were coregistered onto corresponding HU-corrected CT images and the PET images were spatially normalized using normalization parameters of each HU-corrected CT image onto MNI space by the created brain CT template (Fig. 1c). Using the normalized FMM and FBB PET images with WC mask as the reference region, CT-based SUVR parametric PET images were created. The SUVR PET images were used to generate a FMM-FBB global CTX VOI in the same manner as described above for the MRI-based FMM-FBB method (Fig. 1d2). In addition, individual dcSUVR values were calculated using the CT-based FMM-FBB global CTX VOI. CT-based regional VOIs were also de ned by overlapping CT-based FMM-FBB global CTX VOI and AAL atlas (Fig. 1e2). The sub-regions of AAL in the CT-based global CTX VOI were merged into six regions (frontal, PC, parietal, striatum, occipital, and temporal). Regional dcSUVR values were calculated using the six regional VOIs.

Development of MRI and CT-based global and regional dcCL
The dcCL method was used to derive equations from global and regional dcSUVR values using created FMM-FBB VOIs for direct conversion [6]. Fig. 1f1and 1f2 show the summary of methods; each method shows the process of regression equations derived from dcSUVR and dcCL of MRI-based and CT-based methods globally and in six regions, respectively.
The FMM-FBB VOIs of the CT-based method were applied to FMM and FBB PET to acquire dcSUVRs and FMM-FBB VOIs of MRI-based method were used to validate the CT-based method. First, the equations of dcCL conversion from MRI and CT-based dcSUVR and dcCL values were derived using the CL formula globally and in six regions. Second, dcCL scales of MRI and CT-based methods were calculated using the dcCL conversion equations globally and for the six regions.
Validation of the clinical e cacy of CT-based regional dcCLs in the independent cohort To validate the clinical e cacy of CT-based regional dcCLs, 2,245 FMM and FBB PET scans in ADD, aMCI, and cognitive normal (CN) groups were recruited. Gaussian mixture model was performed to determine dcCL cutoffs in 547 NCs 55 years of age or older. Global and six regional cutoffs including frontal, PC, parietal, striatum, occipital, and temporal, were determined using a machine learning technique and the cutoffs were 18.96, 22.32, 28.06, 21.57, 27.02, 25.57, and 23.07, respectively. The group was classi ed into four groups based on global and striatal dcCL cutoffs. First, based on global Aβ dcCL scales, the cohort was classi ed as global (-) and global (+). The global (-) group was further classi ed into regional (-) and regional (+) groups based on regional cutoffs for at least one or more regions. In addition, the global (+) group was further classi ed into striatal (-) and striatal (+) groups based on striatal cutoffs. Thus, the cohort was classi ed into four groups: global (-) and regional (-) Aβ: G(-)R(-); global (-) and regional (+

Statistical analysis
In the head-to-head cohort, group difference and ROC analysis were performed between regional visual positivity and MRI-and CT-based regional dcCL scales in order to validate regional dcCL scales. Regression analysis was performed for reliability between FMM and FBB PET ligands or between MRIbased and CT-based methods using dcSUVR and dcCL scales globally and regionally. The regression was also performed to derive global and regional dcCL formulas from the head-to-head cohort. For precision, the differences in dcCL scales between FMM and FBB ligands or between MRI-based and CT-based methods were investigated using Bland-Altman plots [19]. The absolute value differences between dcCL scales of MRI-based and CT-based methods or between dcCL scales derived based on FMM and FBB ligands were compared using a generalized estimating equation (GEE).
In an independent cohort for clinical validation, the chi-square test for categorical variables and analysis of covariance (ANCOVA) for continuous variables were used to compare the demographics and frequency of APOE4 genotype and MMSE scores among the four groups. To investigate the neuropsychological results among the four groups, ANCOVA was performed after controlling for age and apolipoprotein E ε4

Results
Demographics of the participants Table 1 shows the demographic information of the participants in the regional dcCL development cohort. The average age (SD) of all 63 participants was 58.7 (19.4) years and 58.7% were females. The frequency of APOE-ε4 carriers was 39.7%.

G(-)R(-) global (-) and regional (-) Aβ dcCL scales, G(-)R(+) global (-) and regional (+) Aβ dcCL scales, G(+)Str(-) global (+) and striatal (-) Aβ dcCL scales, G(+)Str(+) global (+) and striatal (+) Aβ dcCL scales
Visual assessment and regional dcCL scales Most of regional MRI-and CT-based dcCL scales showed high area under curve (AUC) values than 0.9 for ve regions classi ed by visual read in Table 2. Total ve regions are signi cant different for both FMM and FBB PET (p<0.001, Additional le1, Figure S1). CT-based global and regional dcCLs The spatial distribution of Aβ deposition in the CT-based CTX VOI did not differ between FMM and FBB as shown in Additional le 1, Figure S2d and 2e. The CT-based CTX VOI regions (Additional le 1, Figure S2f) generally overlapped with MRI-based CTX VOI regions (Additional le 1, Figure S2c). The FMM and FBB SUVR values of head-to-head cohort showed excellent linear correlation and all R 2 values were 0.97 in global and six regional VOIs (Additional le 1, Figure S3b).
The regression equations to convert the FMM dcSUVR into FMM dcCL (Additional le 1, Figure S4c) and FBB dcSUVR into FBB dcCL (Additional le 1, Figure S4d) globally and regionally were calculated. The dcCL scales between FMM and FBB were highly correlated globally and regionally (R 2 = 0.97; Additional le 1, Figure S5b) using the direct comparison of the CT-based method.
The results of MRI-based global and regional dcCLs are described in Additional le 1, Supplementary Methods 3.
Reliability and precision in the MRI-based and CT-based global and regional dcCLs Both FMM and FBB showed SUVRs between MRI-based and CT-based methods were highly correlated globally and regionally in the head-to-dead dataset ( Fig. 2 and Additional le 1, Figure S6a; R 2 = 0.97-0.99). For reliability, the correlation between MRI-based and CT-based global and regional dcCLs was investigated. FMM and FBB showed global and regional dcCLs between MRI-based and CT-based methods were highly correlated (R 2 = 0.97-0.99; Fig. 3 and Additional le 1, Figure S6b).
For precision, the global and regional absolute difference of FMM and FBB dcCLs between MRI-based and CT-based methods were investigated as shown in Additional le 1, Table S1. Fig. 4 shows the distribution plots of absolute FMM dcCLs difference between MRI-based and CT-based methods, which was similar with FBB (Additional le 1, Figure S6c). The GEE results for the absolute difference of dcCLs between FMM and FBB globally and in six regions showed the CT-based method was not different compared with the MRI-based method, globally and regionally (Additional le 1, Table S1 and Additional le 1, Figure S7).

Validation of CT-based global and regional dcCLs in independent participants
In our clinical validation study, the average age (SD) of all participants was 69.6 (10.3) years and 56.1% were females (Table 1). In addition, 32.1% of participants were in the G(-)R(-) group, 11.4% in the G(-)R(+) group, 5.3% in the G(+)Str(-) group, and 52.3% in the G(+)Str(+) group. Figure 5a shows the global dcCL scales in each group. In each group, global dcCL scales tended to increase from G(-)R(-) to G(+)Str(+) group. The global dcCL and SVLT delayed recall was lower in the G(-)R(+) group than in the G(-)R(-) group and CDR-SOB was higher in the G(-)R(-) group than in the G(-)R(+) group (p < 0.05). The global dcCL and SVLT delayed recall were also lower in the G(+)Str(+) group than in the G(+)Str(-) group and CDR-SOB was higher in the G(+)Str(+) group than in the G(+)Str(-) group (p < 0.05) as shown in Figure 5b-e. The number of ADD patients increased per group order is shown in Fig. 5f.

Discussion
In the present study, CT-based regional dcCL scales of Aβ ligands were developed using head-to-head datasets of FMM and FBB PET ligands and their clinical e cacy validated using an independent largesized cohort. The major ndings were as follows: both MRI-based and CT-based dcCL scales showed FMM and FBB were highly correlated globally and regionally; both FMM and FBB showed CT-based regional dcCL scales were mutually highly correlated with MRI-based regional dcCL scales and absolute differences in regional dcCL scales between CT-based and MRI-based methods appeared relatively insigni cant; the G(-)R(+) and G(+)Str(+) groups predicted worse neuropsychological performance than the G(-)R(-) and the G(+)Str(-) groups. Collectively, the results indicate converting FMM or FBB dcSUVR values into the dcCL scales regionally without additional MRI scans is feasible, which could be more accessible to researchers than other approaches and applicable to a variety of different conditions.
The rst major nding was that both MRI-based and CT-based regional dcCL scales showed FBB and FMM were mutually highly correlated, globally and regionally, indicating the CT-based method can harmonize FMM and FBB ligands without paired MRI data. In a previous study by Lilja et al., a PET-only normalization method was developed to resolve variability of Aβ uptake in FMM PET [20]. This study was proposed because MRI images might not always be available in the analysis of Aβ PET uptake. Furthermore, in a recent study from the AIBL group, global CL scales were developed using PET template without MRI images [21]. However, the method developed in the present study allows regional as well as global dcCL scales to be provided in the six regions including striatum. To the best of our knowledge, this is the rst study in which the CT-based regional dcCL scales have been developed.
Secondly, in terms of reliability, both FMM and FBB showed CT-based regional dcCL scales were highly correlated with MRI-based regional dcCL scales. In particular, all regional VOIs showed CT-based dcCL values were highly correlated with MRI-based regional dcCL values. In terms of precision, absolute differences in regional dcCL scales between CT-based and MRI-based methods appeared relatively insigni cant; most of the absolute differences applied to the head-to-head cohort in all ligands were located within the signi cant lines of the Bland-Altman graphs. Furthermore, because absolute differences in CT-based regional dcCL scales between FMM and FBB were not larger than in MRI-based methods, our CT-based regional dcCL scale is a reasonable method to convert regional FMM or FBB dcSUVRs into regional dcCL scales, at least in environments where MRI data are not available and regional Aβ uptake information is needed.
The nal major nding in the clinical validation of the CT-based regional dcCL scales was the G(-)R(+) group had worse neuropsychological performance than the G(-)R(-) group. This nding was consistent with a previous study in AD-related cognitive impairment (ADCI) patients from our group, showing subjects with regionally increased Aβ uptakes had worse memory and hippocampal changes than subjects without regionally increased Aβ uptakes [8]. Furthermore, in the present study, the G(+)Str(+) group had worse neuropsychological performance than the G(+)Str(-) group, which is in agreement with a previous study in 2018 from our group showing the striatal involvement of Aβ is a predictor of poor prognosis. Previously, researchers were unable to detect subjects with regionally increased Aβ uptakes except in individuals with subthreshold global amyloid levels or striatal involvement of Aβ because the conventional dcCL scale only provided information regarding global Aβ levels, not regional Aβ levels. Therefore, CT-based regional dcCL scales used in the present study might provide clinicians with more sensitive diagnostic and prognostic data.
One strength of the study was that a prospectively well-designed head-to-head dataset of FMM and FBB PET ligands was used to develop CT-based regional dcCL scales of Aβ ligands. Another strength was that clinical e cacy of the CT-based regional dcCL scales was validated in an independent large-sized carefully phenotyped cohort using non-invasive amyloid imaging and neuropsychological performance.

Limitations
The present study had several limitations. First, the standard for the presence or absence of Aβ uptakes in each brain region has not yet been established. Thus, further regional pathologic veri cations are needed for more realistic cutoffs. Second, because FBB is derived from Congo red [22] and FMM is based on the chemical structure of thio avin T [23], these two ligands possibly show different dynamic ranges or uptake in the cortex, striatum, and white matter. Although both ligands are comparable for imaging AD pathology in vivo, FMM might be better than FBB for detecting amyloid burden in the striatum. Finally, a partial volume correction (PVC) was not applied to our dataset. Our method still provided information of regional dcCL scales of Aβ ligands without MRI data, which may provide clinicians with a better understanding of biomarker-guided diagnosis and prediction of prognosis.

Conclusions
In conclusion, we were able to nd that our CT-based regional dcCL scales method was comparable to the MRI-based method. Furthermore, unlike the conventional CL method, the information of our method was able to better predict poor clinical impairments. Plots of correlation of global and regional dcSUVR between MRI-based and CT-based methods for FMM.