Evaluation of two clinical Centiloid analysis products for ¹⁸F‐florbetapir PET in cognitively unimpaired elders

Abstract

INTRODUCTION

Recently two software products (MIMneuro and Syngo.PET) received United States Food and Drug Administration (FDA) clearance for Centiloid analysis, thereby introducing Centiloid scoring to the clinical environment. This study compares Centiloid scores by these clinical products and conventional research methods.

METHODS

¹⁸F‐Florbetapir amyloid positron emission tomography (PET) scans (N = 252) of cognitively unimpaired elders were processed using both products and using research pipelines based on the Standard Centiloid method (magnetic resonance imaging [MRI]‐based spatial normalization).

RESULTS

Centiloid scores from both products were highly linearly correlated with each other (r²  = 0.942) and with an MRI‐based research pipeline having matching regions of interest (MIMneuro: r²  = 0.942, Syngo.PET: r²  = 0.971). However, Centiloid scores often differed by more than 10 Centiloids [CL] between methods, with Centiloid calibrations contributing to their inconsistency.

DISCUSSION

Although Centiloid analyses were highly correlated, there was considerable variability in individual cases. Other factors not investigated (e.g., radiopharmaceutical, uptake time) could further contribute to variability. Clinicians should use caution when considering thresholds for amyloid positivity in individual cases.

Highlights

• Two newly available clinical software products, which provide rapid and simple Centiloid scoring, are compared to research processing involving both PET and MRI images.

• Excellent linear correlations were found among the methods.

• Centiloid scores for individual cases often differed by more than 10CL, which should be taken into consideration in clinical reporting.

1. INTRODUCTION

The Centiloid scale was developed with the goal of standardizing quantification of amyloid burden in brain positron emission tomography (PET) scans. ¹ Due to its utility and simplicity, Centiloid scoring has seen great success in the research setting. The potential value of Centiloid scores in the clinical environment is well recognized. ² , ³ Multiple clinical software products provide quantitative measurements of amyloid brain images, typically expressed as standardized uptake value ratio (SUVR) relative to a reference region and as z‐scores with respect to a normal database. ⁴ , ⁵

Two new releases of clinical software products include Centiloid scaling along with SUVR and z‐scores: MIMneuro (MIM Software) and Syngo.PET Cortical Analysis (Siemens Healthineers). Both recently received United States Food and Drug Administration (FDA) clearance for Centiloid analysis in 2024. Unlike Standard Centiloid processing, ¹ which spatially normalizes a high‐resolution MRI to a standard template and applies this normalization to co‐registered PET images, these software products perform spatial normalization based on PET images alone, since a high‐resolution MRI often is not clinically available. The software developers used available datasets to calibrate and validate their Centiloid scaling.

Consistency among research Centiloid pipelines has been well studied, ⁶ , ⁷ , ⁸ but not for these clinical Centiloid products. To evaluate the consistency and variability of Centiloid processing, a large research dataset was processed using these software products and research pipelines based on the Standard Centiloid method.

RESEARCH IN CONTEXT

1. Systematic review: PubMed was used to review the literature. Numerous publications describe Centiloid analysis using research processing methods. Two commercial software programs (MIMneuro and Syngo.PET) providing rapid Centiloid analysis were recently cleared by the FDA. Direct comparisons to research pipelines based on the Standard Centiloid method have not yet been published.

2. Interpretation: Using 252 ¹⁸F‐florbetapir scans of cognitively unimpaired elders, Centiloid scores from both clinical software products were highly linearly correlated with versus research pipelines. However, Centiloid scores often differed by greater than 10CL, suggesting caution if considering cutoff scores in individual cases.

3. Future directions: Other factors not studied here contribute to Centiloid variability, including various radiopharmaceuticals, scanners, and image reconstruction settings. It is anticipated that other laboratories will publish their own comparisons of Centiloid scores by these and other future software products. This future work will further inform clinicians of the accuracy and utility of Centiloid scoring.

2. METHODS

PET ¹⁸F‐florbetapir and MRI images had been acquired for the IMMUNE‐AD study (NIH R01‐AG022304). This cohort consists of cognitively unimpaired elders (age 65–80) with varying degrees of genetic risk (half were apolipoprotein E [APOE] ε4 carriers) and having both PET and MRI at baseline (N = 139) and at 2 years (N = 113). PET studies (mCT 4R PET/CT, Siemens) were acquired and reconstructed following the SCAN protocol. ⁹ A 3D Gaussian postfilter (5 mm full‐width at half‐maximum [FWHM]) was applied to harmonize spatial resolution with typical brain PET scans. ¹⁰ MRI T1 MPRAGE images (Prisma 3T, Siemens) were acquired with 1 mm voxels. Centiloid was calculated by the following four methods.

Method 1: Standard Processing. Following the method of Klunk et al ¹ and using SPM8 software, MRI images were spatially normalized to the MNI152 template. PET images 50–60 min post‐injection were registered to the MRI images, summed, and transformed according to the MRI's spatial normalization. PET SUVR was calculated using the Standard Centiloid regions ¹ and Avid (Clark) regions ¹¹ available from GAAIN. ¹² The local processing pipeline was validated by replication analysis of GAAIN datasets for Pittsburgh Compound B (¹¹C‐PIB) (r ² = 0.999, slope = 1.005, intercept = 0.31) and for ¹⁸F‐florbetapir (r ² = 0.993, slope = 0.987, intercept = −0.52) using the published calibration for Standard regions. ¹³

Method 2: Avid/Standard: Both clinical products utilize Clark Atlas regions (Avid volumes of interest [VOIs], available from GAAIN), which differ from Standard Centiloid regions. For improved consistency, the Standard Processing pipeline was adapted to use the Avid regions for PET SUVR (average over six cortical regions). A linear fit of ¹⁸F‐florbetapir SUVR (Avid regions) to PIB SUVR (Standard regions) yielded the Centiloid calibration for ¹⁸F‐florbetapir with Avid regions and Standard processing (“Avid/Standard”): 176.89 × SUVR – 166.85.

Unlike the above MRI‐based pipelines, the clinical products MIMneuro and Syngo.PET perform spatial normalization based on PET images alone. The programs run semi‐automatically, requiring the user to inspect the PET image alignment to the template. If needed, the user manually adjusts the seed image and restarts the automated spatial normalization. Further manual adjustment of the cortical and reference regions was deemed unecessary and was not performed.

Method 3: MIMneuro. Using version 7.3.2, PET images 50–60 min post‐injection were coregistered, summed, and reoriented along standard brain axes, then processed using the Neuro Amyloid – Centiloid Analysis workflow.

Method 4: Syngo.PET. The identical pre‐processed PET images from Method 3 were used. Using version VE70, these images were processed using the Neuro Cortical Analysis workflow.

Linear correlations were evaluated between the four Centiloid processing methods. Quantitative consistency in Centiloid scores was evaluated based on residue analysis (Bland–Altman plots).

3. RESULTS

Both software products were practical and straightforward to use. Workflow execution was under 1 min and nearly always automatic. Manual restarting of spatial normalization was required only for 2 (MIMneuro) or 5 (Syngo.PET) of 252 studies. Successful spatial normalization was confirmed visually for all studies.

Centiloids by Avid/Standard and Standard Processing were highly linearly correlated (coefficient of determination: r²  = 0.977). Standard deviation of residues (σ) was 4.2CL, which reflects the variability associated with Avid VOIs versus Standard VOIs.

Very good linear correlation versus Standard Processing was observed for MIMneuro (r²  = 0.932, σ = 8.0CL) and Syngo.PET (r²  = 0.954, σ = 7.7CL). Variability is attributed to differences in regions and spatial normalization.

MIMneuro and Syngo.PET Centiloids were both highly correlated versus Avid/Standard (Figure 1A,B), with r²  = 0.942 and 0.971 and σ = 7.3CL and 6.3CL, respectively. Since consistent regions were used for all three methods, this variability is attributed to spatial normalization methods (MRI‐based vs. PET‐based).

FIGURE 1.

Correlation graphs of Centiloid measurements. (A) MIMneuro versus Avid/Standard, r²  = 0.942. (B) Syngo.PET versus Avid/Standard, r²  = 0.971. (C) MIMneuro versus Syngo.PET, r²  = 0.942. Linear fit is shown with a solid line, and the line of equality is shown with a dashed line.

Centiloid correlation of MIMneuro versus Syngo.PET (Figure 1C) had r²  = 0.942 and σ = 7.7CL, with variability attributed to differences in their PET‐based spatial normalizations.

Bland–Altman plots (Figure 2A–C) show that Centiloid differences frequently exceed 10CL and occasionally exceed 20CL. Compared to Avid/Standard, absolute differences exceeded 10CL in 29% and 11% of cases for MIMneuro and Syngo.PET, respectively. When compared to each other, the absolute differences between MIMneuro and Syngo.PET exceeded 10CL in 24% of cases.

FIGURE 2.

Bland–Altman plots of Centiloid measurements corresponding to Figure 1. (A) MIMneuro versus Avid/Standard. (B) Syngo.PET versus Avid/Standard. (C) MIMneuro versus Syngo.PET. Solid and dashed horizontal lines correspond to mean ± 2 × (standard deviation). The shaded regions indicate agreement within ± 10CL. The percentage of cases with differences greater than +10CL or less than −10CL are 29%, 11%, and 24%, respectively. CL, Centiloid.

4. DISCUSSION

Clinical software programs intended for routine use must be simple, rapid, robust, and consistent with established research methods. Both programs succeed in this regard. Linear correlations with MRI‐based Standard Processing were quite good (r²  = 0.932 (MIMneuro) and 0.954 (Syngo.PET)) and improved further with Avid/Standard having matching regions and SUVR calculation (r²  = 0.942 and 0.971, respectively).

Both MIMneuro and Syngo.PET perform spatial normalization based on PET images alone. This PET‐based approach also was used by Avid's research processing. ¹⁴ SUVR measurements between MIMneuro and Avid's research processing were reported to be highly correlated (r²  = 0.98). ¹⁵ Even so, the loss of accuracy with respect to MRI‐based methods must be considered. ¹⁶ Here, Avid/Standard uses regions consistent with the clinical software programs, focusing on spatial normalization differences. Syngo.PET correlated highly with Avid/Standard (r²  = 0.971), demonstrating excellent consistency with MRI‐based spatial normalization. MIMneuro correlated very well (r²  = 0.942) but less so than Syngo.PET (p < 0.01).

A second contributor to accuracy is Centiloid calibration. All four pipelines’ Centiloid calibrations are based directly or indirectly on the available GAAIN ¹⁸F‐florbetapir dataset contributed by Navitsky et al. ¹³ One then would expect these Centiloid calibrations to be consistent. However, the slopes and intercepts of the linear correlations (Figure 1A–C) differ noticeably from their ideal values of 1 and 0, thereby contributing non‐statistical error to Centiloid. When combined with statistical uncertainties and considering outliers, Centiloid differences between pipelines could be significant (Figure 2A–C). Mean Centiloid differences ranged from 3.2CL to 6.6CL. A sizable fraction of cases differed by more than 10CL, with some differing by more than 20CL.

A limitation of this work is that it focused on technical differences in Centiloid processing. Other factors, such as radiopharmaceutical, scanner, protocol, image reconstruction, and scan‐rescan variability, motivate further study. ¹⁶ , ¹⁷ These factors all contribute to Centiloid variability and further limit the utility of thresholds. Another limitation is that the cohort was cognitively unimpaired and scanned under a controlled research protocol. Amyloid PET currently is not indicated for the cognitively unimpaired. Although this cohort is not representative of the clinical population, the data are enriched with borderline‐abnormal cases for which Centiloid accuracy is of particular interest. Centiloid variability as scanned under clinical protocols deserves further study. For example, the uptake time was strictly specified; however, the clinical software programs neither specify the uptake time nor correct for variable uptake time. Centiloid variability should be evaluated for a larger range of uptake time (according to package inserts: 30–50 minutes for ¹⁸F‐florbetapir, 45–130 minutes for ¹⁸F‐florbetaben, and 60–120 minutes for ¹⁸F‐flutametamol).

In a clinical environment, one is concerned with the accuracy of individual measurements and how they might influence the management of individual patients. Centiloid has value to clinicians, not only to indicate the degree of amyloid accumulation in amyloid‐positive patients, but also to assist in elucidating amyloid positivity in cases of early disease. ³ , ¹⁸ Thresholds for amyloid positivity are reported in research studies; for example, Navitsky et al reported that a Centiloid threshold of 24.1 discriminated none‐or‐sparse versus moderate‐to‐frequent amyloid plaques. ¹³ It is important to note that this Centiloid threshold is specific to the Avid research pipeline used in that study. Ideally, the Centiloid calibration method harmonizes differences between processing pipelines, scanners, and radiopharmaceuticals, and a published threshold would still be valid. However, this requires consistent Centiloid calibrations, which differed among the pipelines investigated here. For example, a linear fit to MIMneuro Centiloid versus Syngo.PET Centiloid had a slope of 0.916 and an intercept of 4.73, significantly worse than expected for Centiloid replication analysis (slope: 1 ± 0.02, intercept: ± 2.0). ¹

Recently, the AMYPAD consortium recommended negativity and positivity thresholds of < 10CL and > 30CL and considered intermediate values related to increasing risk of disease progression. ¹⁹ This suggests a central threshold of 20CL with accuracy of ± 10CL. The observed variability among pipelines (often exceeding 10CL) suggests that clinicians should use caution if considering thresholds in individual cases.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All study procedures were approved by the Cleveland Clinic Institutional Review Board, and each participant provided informed consent prior to participation.

Supporting information

Supporting Information

DiFilippo FP, Rao SM. Evaluation of two clinical Centiloid analysis products for ¹⁸F‐florbetapir PET in cognitively unimpaired elders. Alzheimer's Dement. 2025;17:e70163. 10.1002/dad2.70163