AI-enhanced Centiloid quantification of amyloid PET images

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomarkers

Background:

• The Centiloid scale is standard for amyloid PET quantification but faces challenges with tracer and scanner variability.
• Accurate amyloid quantification is crucial for research and clinical applications in Alzheimer's disease.

Purpose of the Study:

• To introduce DeepSUVR, a novel deep learning method for correcting Centiloid quantification in amyloid PET scans.
• To improve the consistency and reliability of amyloid PET data across different tracers, scanners, and studies.

Main Methods:

• DeepSUVR utilizes a deep learning approach that penalizes implausible longitudinal trajectories during training.
• The model was trained on a large dataset (7,149 scans) from AIBL/ADNI and validated on 15,807 scans from 10 external datasets.
• The method learns standardized uptake value ratios (SUVR) correction factors without needing longitudinal data at inference.

Main Results:

• DeepSUVR increased tracer correlation and reduced variability in amyloid-negative individuals.
• It demonstrated stronger associations with cognitive performance, visual reads, and neuropathology.
• The method significantly improved longitudinal consistency and increased the effect size for detecting treatment-related amyloid changes.

Conclusions:

• DeepSUVR substantially advances amyloid PET quantification, outperforming standard methods.
• This AI-driven approach is vital for clinical decision-making and detecting early or subtle amyloid accumulation.
• DeepSUVR offers a practical alternative for integration into existing PET imaging pipelines.