Harmonizing florbetapir and PiB PET measurements of cortical Aβ plaque burden using multiple regions-of-interest and machine learning techniques: An alternative to the Centiloid approach

Abstract

Introduction: Machine learning (ML) can optimize amyloid (Aβ) comparability among positron emission tomography (PET) radiotracers. Using multi-regional florbetapir (FBP) measures and ML, we report better Pittsburgh compound-B (PiB)/FBP harmonization of mean-cortical Aβ (mcAβ) than Centiloid.

Methods: PiB-FBP pairs from 92 subjects in www.oasis-brains.org and 46 in www.gaain.org/centiloid-project were used as the training/testing sets. FreeSurfer-extracted FBP multi-regional Aβ and actual PiB mcAβ in the training set were used to train ML models generating synthetic PiB mcAβ. The correlation coefficient (R) between the synthetic/actual PiB mcAβ in the testing set was assessed.

Results: In the testing set, the synthetic/actual PiB mcAβ correlation R = 0.985 (R² = 0.970) using artificial neural network was significantly higher (p ≤ 6.6e-4) than the FBP/PiB correlation R = 0.927 (R² = 0.860), improving total variance percentage (R² ) from 86% to 97%. Other ML models such as partial least square, ensemble, and relevance vector regressions also improved R (p = 9.677e^-05 /0.045/0.0017).

Discussion: ML improved mcAβ comparability. Additional studies are needed for the generalizability to other amyloid tracers, and to tau PET. Highlights Centiloid is a calibration of the amyloid scale, not harmonization. Centiloid unifies the amyloid scale without improving inter-tracer association (R² ). Machine learning (ML) can harmonize the amyloid scale by improving R² . ML harmonization maps multi-regional florbetapir SUVRs to PiB mean-cortical SUVR. Artificial neural network ML increases Centiloid R² from 86% to 97%.

Keywords: Centiloid; PiB; amyloid PET harmonization; artificial neural network; florbetapir; machine learning.