Abstract
As Centiloids are increasingly used in trials and clinical settings to quantify amyloid-β (Aβ) PET, better characterization of sources of measurement error is essential. We examined 2 potential factors driving it: variability in the estimated coefficients in the SUV ratio-to-Centiloid conversion equation related to random sampling of the calibration dataset and PET image resolution. Methods: First, we analyzed [(11)C]PiB scans in 200 participants with a clinical diagnosis of Alzheimer disease (cAD) and 114 Aβ-negative participants. PET scans were processed using the standard Centiloid pipeline and a nonstandard MRI-based pipeline (native space, cerebellar cortex as reference). We split data into training and test datasets (n = 157 each) to compare conversion equations in subsamples with an n of 10-30 Aβ-negative and 15-50 cAD participants. Second, all [(11)C]PiB images, along with 604 [(18)F]florbetaben and 538 [(18)F]florbetapir images, were reduced from high (6/7 mm(3)) to medium (8 mm(3)) and low (10 mm(3)) resolution and resulting Centiloids were compared between resolutions. rPOP and CapAIBL, 2 PET-only processing pipelines, were used to explore the effects of the PET spatial resolution across different pipelines. Results: In the smallest required sample of 15 cAD and 10 Aβ-negative participants, conversion error was 1.7 Centiloids at 25 Centiloids and 3.4 Centiloids at 100 Centiloids. Error decreased to 1.0 Centiloid at 25 Centiloids and 2.0 Centiloids at 100 Centiloids, when including 50 cAD and 30 Aβ-negative participants. Lower image resolution was associated with a systematic difference in Centiloids, especially in highly positive [(11)C]PiB scans: a scan estimated as 100 Centiloids in high resolution was quantified as 94.2 Centiloids and 84.9 Centiloids at medium and low resolution. When a [(11)C]PiB scan was quantified as 25 Centiloids in its high resolution, lowering its resolution resulted in reduced values of 23.5 Centiloids and 20.6 Centiloids for medium and low resolution, respectively. Similar trends were observed for [(18)F]florbetaben and [(18)F]florbetapir scans. Conclusion: A relatively accurate SUV ratio-to-Centiloid conversion equation for level 2 analyses can still be achieved with a minimally required datasets. Increasing the number of cAD participants reduced error at higher values, whereas adding Aβ-negative participants had little effect. Image resolution significantly impacts Centiloids in highly positive scans and should be considered when interpreting data acquired with different settings. Errors remain minimal at 25 Centiloids, the typical cutoff for determining Aβ positivity.