Biomarkers

Centiloid quantification relies on a Standardised Uptake Value Ratio (SUVR) that could be subject to noise, spill-in, and specific binding in the reference region. We developed a novel deep learning method that learns scan-specific variability from longitudinal trends to correct SUVR quantification and then validated it in 7 large cohorts.

2,281 participants with 2+ visits in AIBL/ADNI (7,380 scans) had their Amyloid PET images spatially normalised and quantified using the Centiloid SPM8 pipeline. A deep learning network (DeepSUVR) was trained to predict a SUVR correction factor for each spatially normalised image, by penalising unexpected temporal changes based on the deviation from the Centiloid/Year vs mean Centiloid curve. Importantly, while the model requires longitudinal data for training, inference is performed on each visit independently. The model was trained using 5-fold cross-validation on ADNI+AIBL and evaluated on the out-of-fold AIBL+ADNI as well as OASIS3, AMYPAD, MCSA, HABS-HD and A4-LEARN studies (8,806 participants, 12,320 scans), comparing DeepSUVR with standard CL and when using a composite WCb+WM reference region for ¹⁸F-Florbetapir.

Fitting a bimodal Gaussian mixture on the baseline CL of each study, DeepSUVR improved the alignment of the first peak across all studies, as well as reducing their standard deviations (Figure 1A) with similar findings across tracers (Figure 1B). DeepSUVR CL had the strongest correlation between baseline CL and MMSE, best group separation between CDR 0/0.5/1 (Table 1A) and highest AUC against visual reads (Table 1B). DeepSUVR improved model fitness (Spearman rank) and inter-study model agreement in the longitudinal trajectories across studies in both training (Figure 2A) and testing cohorts (Figure 2B). DeepSUVR increased the effect size for the reduced rate of CL increase per year seen with treatment in the A4 study, while reducing the variability in the Amyloid negative subjects (Figure 2C).

Deep learning provides a significant advance in PET quantification of amyloid, outperforming standard methods both cross-sectionally and longitudinally in both observational and interventional studies. DeepSUVR improves the ability to harmonise large datasets and different PET tracers. This is particularly important for consistent decision making and when the outcome measure of an intervention is expected to be subtle such as altering the rate of accumulation.