Revisiting Centiloids using AI

Abstract

The Centiloid scale is the standard for Amyloid $\left(\text{A}\beta \right)$ PET quantification, widely used in research, clinical settings, and trial stratification. However, variability between tracers and scanners remains a challenge. This study introduces DeepSUVR, a deep learning method to correct Centiloid quantification, by penalising implausible longitudinal trajectories during training. The model was trained using data from 2,098 participants (6,762 $\text{A}\beta$ PET scans) in AIBL/ADNI and validated using 15,806 $\text{A}\beta$ PET scans from 10,543 participants across 10 external datasets. DeepSUVR increased correlation between tracers, and reduced variability in the $\text{A}\beta$ -negatives. It showed the strongest association with cognition, highest AUC against visual reads and best longitudinal consistency between studies. DeepSUVR also increased the effect size for detecting lower Centiloid increase per year in the A4 study. DeepSUVR advances $\text{A}\beta$ PET quantification, outperforming standard approaches, which is particularly important for consistent decision making and to detect subtle and early changes in clinical interventions.

Figure 1

Scatterplots comparing PiB vs FBP in the OASIS Head-to-Head (H2H) dataset (left) and each ¹⁸F-Tracer versus PIB in the GAAIN Head-to-Head Calibration dataset (Right), using ${\text{CL}}_{\text{Std}}$ (top) and ${\text{CL}}_{\text{DS}}$ (bottom). The correlation between each pair of tracers is assessed using the coefficient of determination ${\text{R}}^2$. The standard deviation in the Young Controls (YC) for each tracer is denoted as $\left.\sigma ({\text{Tracer}}_{\text{YC}}\right)$. Note that for FMM, one YC with a CL>20 on both PIB and FMM across all methods was determined to be an outlier and excluded from the standard deviation calculation.

Figure 2

Histogram distribution of Centiloid values across the 12 cohorts using the Standard and DeepSUVR CL methods. For each approach, a Gaussian mixture is fitted to the distribution of Centiloid values of each study (top) and each tracer (bottom), and the average mean [min,max] and average standard deviation [min,max] of the first peak across all studies (tracer respectively) is reported. The dashed lines mark the 0CL and 20CL. Significantly lower variability in the means and standard deviations when comparing ${\text{CL}}_{\text{DS}}$ to ${\text{CL}}_{\text{Std}}$ across studies/tracers based on bootstrapping are indicated using: *** (p<0.001) for lower variabilities in the means and ### (p<0.001) for lower variabilities in the standard deviations.

Figure 3

Rate of CL change per year compared to mean CL computed from the 2 training (AIBL/ADNI) cohorts (a.), the 7 testing cohorts with longitudinal data (A4-Placebo, AMYPAD, HABS-HD, OASIS, DLBS, MCSA, WRAP) (b.) and the placebo and treatment arms of the A4 Study (c.) using the Standard (left) and DeepSUVR CL (right) methods. Each point represents the mean and rate of change between a pair of consecutive visits from the same participant. Each curve shows a 5th order polynomial fitted to each study. The Spearman rank correlation between Mean CL and CL/Year is denoted using r (a. and b.), while the effect size of the CL accumulation per year between the 2 arms of the A4 study is denoted using Cohen’s d (c.). The vertical dashed lines mark the 0CL and 20CL, while the horizontal dashed line mark the 0CL/Year. Significantly higher correlation or effect size when comparing ${\text{CL}}_{\text{DS}}$ to ${\text{CL}}_{\text{Std}}$ across studies/tracers based on bootstrapping are indicated using: *** p<0.001.

Figure 4

Axial views of the Standard CL reference mask (row1) and the new reference mask derived from DeepSUVR (row2), the Standard CL target mask (row3) and the new target mask derived from DeepSUVR (row4)

Figure 5

Framework used for training (top) and inference (bottom) using DeepSUVR. While training requires pairs of images from the same participant, inference is run on single images.

Figure 6

The penalties in the loss functions used to constrain the model: for each pairs of TPs, penalise the Centiloid decreasing over time (a); penalise the distance to the population curve of Mean CL vs Rate of change (b); for each batch, penalise the regression line of the standard CL vs corrected CL when it deviates from slope=1 and intercept=0.