Centiloid scaling for quantification of brain amyloid with [¹⁸F]flutemetamol using multiple processing methods

Alzheimer’s disease (AD) is characterised by two key pathological findings, β-amyloid plaques and neurofibrillary tangles [1]. As possibly the earliest pathology, β-amyloid is a compelling therapeutic target [2]. The use of imaging biomarkers to visualise and measure the β-amyloid plaque load in individuals was introduced by Klunk et al., where a ¹¹C-labelled Thioflavin-T-based molecule was developed to visualise amyloid plaque presence in vivo using PET imaging [3, 4]. Although [¹¹C]PiB (PiB, Pittsburgh Compound-B) became the ‘Gold-Standard’ amyloid PET tracer for research studies, its use is limited by the short half-life (20 min) of ¹¹C, requiring an on-site cyclotron when imaging with this tracer. This prompted the generation and clinical approval of ¹⁸F-labelled tracers (110 min half-life), allowing greater distribution and utilisation in PET centres [5]. Three tracers, [¹⁸F]florbetapir (Amyvid™), [¹⁸F]florbetaben (NeuraCeq™) and [¹⁸F]flutemetamol (Vizamyl™), are validated with post-mortem studies and have been approved by regulatory authorities [6‐8]. Another tracer, [¹⁸F]NAV4694, has also been studied in a limited setting [9]. These tracers all show increased cortical retention in AD subjects. Each tracer has its own unique set of cortical and reference regions, methods for evaluation and positive PET cut-off points associated with quantitative use. In addition, differences in their dynamic range, kinetics and non-amyloid white matter binding all add up to make comparing data sets across studies, and across groups scanned with more than one of these tracers, complex.

Consequently, there is a need to standardise the methods for data collection and analysis to better aid cross-centre, multi-tracer utility. Standardised units would also allow better interpretation of longitudinal changes and improve how sites monitor disease progression and whether any potential therapeutic effects are observed.

The Centiloid Project was initiated to derive a standardised quantitative amyloid imaging measurement scale, based upon normalisation of data from the ¹⁸F-tracers to that of PiB. In this linear scale, young controls (≤ 45 years) have a mean of zero Centiloid units (CL) and typical mild to moderate AD patients score on average 100 CL [10]. The data set used to determine PiB Centiloids is freely available on the Global Alzheimer Association Interactive Network website (GAAIN; http://​www.​gaain.​org), together with standardised cortical and whole cerebellum volume of interest (VOI) templates.

Standardised scaling, using Centiloids, has been recently reported for other amyloid tracers, [¹⁸F]florbetaben, [¹⁸F]NAV4694 and [¹⁸F]florbetapir, allowing comparison of these tracers with PiB [11‐13].

This work reports the application of the Centiloid scaling methods to images obtained using [¹⁸F]flutemetamol (Vizamyl™, GE Healthcare). We also report the utility of alternate processing pipelines and assess the robustness and margin of error of the Centiloid analysis system through test-retest analysis on the standard and other processing systems.

Validation of our local standard Centiloid (SPM8) process pipeline using the GAAIN data gave an excellent correlation, local SPM8 Centiloids = 1.00 × GAAIN Centiloids – 0.07 (R² = 0.999), compared with the published data. This falls well within the minimum specified acceptance criteria defined by Klunk et al.

A standardised, robust method of quantification is key to utilising amyloid PET in disease prediction, diagnosis and progression. The methods described by Klunk et al. have provided the framework to allow sites to compare multiple tracers using a single scale, Centiloids. Essential to this method of analysis is the accurate establishment and calibration of the processing pipeline. Here, we have described and defined the conversion of [¹⁸F]flutemetamol SUVR derived using the standard Centiloid methods (SPM8) to Centiloid units. SPM8 was the recommended image processing platform described in the methods of Klunk et al. and was selected by us when replicating the process pipeline. More recent versions of SPM have been introduced since, with the use of SPM12 reported by other groups [21]. However, we chose to use SPM8 to ensure we followed the process as accurately as possible. We also investigated the utility of two alternate image processing software platforms (PMOD and FSL) to replicate the Centiloid process and generate [¹⁸F]flutemetamol results. These platforms were selected as they were available and routinely used in-house.

Whole cerebellum was selected as the reference region for use with Centiloid processing. This differs from the preferred reference region, cerebellar cortex, recommended for [¹⁸F]flutemetamol. Again, whole cerebellum was used in this study in order to ensure the accurate replication of the Centiloid process. The use of alternate reference regions may be investigated in future work.

The data show that [¹⁸F]flutemetamol is a suitable tracer for Centiloid conversion, using the standard SPM8 process. There was strong correlation between [¹⁸F]flutemetamol and PiB, indicating that the tracers show similar uptake and kinetics. This was also apparent from the low variance values observed for the YHC subjects. When compared with other tracers, [¹⁸F]flutemetamol performed favourably when converted to Centiloids, with a better variance ratio (1.19) than that reported for [¹⁸F]florbetaben (1.96) and [¹⁸F]florbetapir (4.62) [11, 13]. This would be expected given the similar chemical structures of PiB and [¹⁸F]flutemetamol.

The utility of PMOD and FSL to process [¹⁸F]flutemetamol images and derive SUVR values, for conversion to Centiloid, was also assessed in this study. Both PMOD and FSL performed well, producing data that was comparable to the standard SPM8 methods. Validation of each pipeline against the GAAIN data again gave excellent correlation when compared with the published data, with both pipelines fulfilling the required acceptance criteria. The equations derived for each method were then transformed to ‘Standard’ SPM8-equivalent Centiloids, allowing a true comparison of values. This shows that Centiloid scaling is robust and can be implemented on a number of platforms, not just the initially recommended SPM8 version. Furthermore, this work provides a straightforward framework for implementation and validation of Centiloid scaling using any other processing pipeline.

Further, the test-retest difference has been assessed and produced an average of approximately of 2% or less between test and retest Centiloid values on each pipeline. The different image processing pipelines did not produce significantly different results for the test-retest data, with an average of ≤ 2% difference between these pipelines: SPM8, PMOD and FSL.

Previously published data for a small cohort of AD subjects found the SUVR test-retest variability in a composite cortical region averaged 1.5% (range 0.9 to 2.4%, n = 5), comparable with our findings [14]. However, SUVR scales have a measurement offset which makes SUVR percentage differences in test-retest look more favourable than those for binding potential (BP), and differences in SUVR-1 test-retest values are a better comparison as BP can be approximated by SUVR-1 [22, 23]. Importantly, the methodology of creating Centiloid values has an in-built subtraction of average healthy control SUVRs to zero, therefore providing a more robust measure than direct SUVR percentage differences that is similar to that for SUVR-1. The low difference in test-retest Centiloid results is comparable with the difference in estimated binding potential related measurements (SUVR-1).

This low difference in the test-retest data, and the favourable variance in young healthy controls means that [¹⁸F]flutemetamol Centiloid are suitable for comparing images, or monitoring brain amyloid levels in therapeutic clinical trials, with a high degree of sensitivity to small changes.

This work will allow us to analyse and compare data from various studies, a key future aim of the Centiloid project. The conversion equation allows us to apply Centiloid scaling to subjects for which [¹⁸F]flutemetamol/MRI scans but no PiB data are available for analysis.

The use of different scanners in this study was not taken in to account when analysing the data. It would be interesting to further investigate this to fully understand any variations and help interpret the results.

Further, a recent study by Bourgeat et al. looked at the utility of an alternate image processing platform, CapAIBL [24]. CapAIBL utilises a PET-only approach, overcoming the requirement for a corresponding MR image, and is particularly useful where MR imaging is not possible. The authors reported similar results for the conversion of [¹⁸F]flutemetamol through this pipeline. The application of such PET-only quantification methods could lead to a readily adopted clinical quantification method as images could be processed directly from the PET scanner. To that end, further work to investigate the utility of a PET-only method based on CortexID (AW Workstation, GE Healthcare), a dedicated platform for reviewing [¹⁸F]flutemetamol images is on-going.

[¹⁸F]flutemetamol data can now be expressed in Centiloid units, enhancing its utility in both clinical and research applications for β-amyloid imaging. Standardised quantification can provide supplementary information to compliment visual assessment, especially in equivocal cases, and also provide a means to assess patients longitudinally. The standard Centiloid method also demonstrates that [¹⁸F]flutemetamol has favourable performance compared with PiB and other β-amyloid tracers. Test-retest difference averaged 2%, with no difference between image processing pipelines. Centiloid scaling is robust and can be implemented on a number of platforms.