Biomarkers.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarker Quantification

Background:

• Centiloid (CL) quantification using Standardised Uptake Value Ratio (SUVR) is susceptible to noise, spill-in, and reference region binding.
• Accurate amyloid PET quantification is crucial for diagnosing and monitoring Alzheimer's disease and evaluating treatment efficacy.

Purpose of the Study:

• To develop and validate a novel deep learning method (DeepSUVR) for correcting SUVR quantification in amyloid PET imaging.
• To assess DeepSUVR's performance in harmonizing large, multi-cohort datasets and improving longitudinal tracking of amyloid accumulation.

Main Methods:

• A deep learning network (DeepSUVR) was trained on longitudinal data from 2,281 participants (7,380 scans) in the ADNI and AIBL cohorts.
• The model learns scan-specific variability to predict SUVR correction factors, penalizing temporal deviations from established trends.
• Validation was performed across seven large cohorts (8,806 participants, 12,320 scans), comparing DeepSUVR with standard Centiloid methods.

Main Results:

• DeepSUVR improved the alignment and reduced variability of baseline Centiloid values across multiple studies and tracers.
• DeepSUVR demonstrated stronger correlations with cognitive scores (MMSE) and better diagnostic group separation (CDR) compared to standard methods.
• The method enhanced longitudinal model agreement and increased the effect size for treatment-induced changes in amyloid accumulation rates.

Conclusions:

• Deep learning, specifically DeepSUVR, represents a significant advancement in amyloid PET quantification, outperforming traditional methods.
• DeepSUVR facilitates harmonization of large datasets and diverse PET tracers, crucial for consistent clinical decision-making and subtle intervention outcome detection.
• This approach holds promise for improving the reliability of amyloid PET in both observational research and clinical trials.