Alzheimer's Imaging Consortium.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarker Quantification

Background:

• Centiloid quantification for amyloid PET relies on Standardised Uptake Value Ratio (SUVR), susceptible to noise and reference region issues.
• A novel deep learning method was developed to address these limitations by learning scan-specific variability.

Purpose of the Study:

• To develop and validate a deep learning approach (DeepSUVR) for correcting SUVR quantification in amyloid PET.
• To improve the accuracy and harmonization of amyloid PET data across multiple cohorts and tracers.

Main Methods:

• Trained a deep learning network (DeepSUVR) on longitudinal data from 2,281 participants (7,380 scans) to predict SUVR correction factors.
• Validated DeepSUVR across 7 large cohorts (8,806 participants, 12,320 scans), comparing it with standard Centiloid methods.

Main Results:

• DeepSUVR improved the alignment and reduced variability of baseline Centiloid values across studies and tracers.
• Demonstrated stronger correlations with cognitive scores (MMSE) and better diagnostic group separation (CDR).
• Enhanced longitudinal model agreement and increased effect size for treatment-related changes in clinical trials.

Conclusions:

• Deep learning significantly advances amyloid PET quantification, outperforming standard methods cross-sectionally and longitudinally.
• DeepSUVR improves data harmonization across diverse datasets and PET tracers.
• This method is crucial for consistent decision-making, especially when detecting subtle intervention effects.