Improved centiloid robustness using non‐negative matrix factorization

• Published in: Alzheimer's & dementia Vol. 16
• Main Authors: Bourgeat, Pierrick, Dore, Vincent, Ames, David, Martins, Ralph N, Masters, Colin L, Rowe, Christopher C, Fripp, Jurgen, Villemagne, Victor LL
• Format: Journal Article
• Language: English
• Published: 01.12.2020
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.040085

Background Centiloid was introduced to harmonise Amyloid PET quantification across different tracers, scanners and analysis techniques. Unfortunately, Centiloid suffers from quantification disparities in longitudinal analysis when switching from different tracers or scanners. In this work, we aim to reduce such variability using a different analysis technique on the existing calibration data. Method All the PET images from the calibration dataset, along with 3658 PET images from the AIBL study were analysed using the recommended SPM pipeline. All the images were SUVR normalised using the whole cerebellum and skull stripped. The nomalised PiB images were decomposed into 6 components using non‐negative matrix factorisation (NMF). NMF is similar to a PCA decomposition but imposes that all components and associated coefficients remain positive. The NMF coefficients related to the first component were strongly correlated with global SUVR (R2 = 0.93), and were subsequently used as a surrogate for Aβ retention. The NMF coefficients were calibrated to Centiloid using the anchoring groups in the PiB calibration dataset. For each tracer of the calibration dataset, a NMF decomposition was computed, where the coefficients of the 1st component were fixed, and set to those of their corresponding PiB, hence forcing the 1st component of the NMF decomposition to match across tracers. All PET images from AIBL were subsequently decomposed using the computed NMF, and their coefficients transformed into Centiloids. Result In AIBL, the correlation between the standard Centiloid and the novel CentiloidNMF was high in each tracer with 0.93 in 11C‐PiB, 0.97 in 18F‐NAV, 0.96 in 18F‐FBB, 0.91 in 18F‐FLUTE and 0.86 in 18F‐AV45. The longitudinal consistency was assessed using average fitting error of a linear regression model applied to each subject with 3 or more timepoints, and also by measuring the number of subjects with longitudinal changes >20CL/year, which are physiologically implausible and likely caused by noise. The average fitting error was decreased by 15%, and the number of subjects with changes >20CL/year was halved from 41 to 20. Conclusion We have proposed an image driven method to measure Centiloids. The method is highly correlated with standard Centiloids and reduces its longitudinal variability when switching scanners and tracers.