Validation of deep learning-based nonspecific estimates for amyloid burden quantification with longitudinal data.

Area of Science:

• Neuroimaging
• Biomarkers
• Artificial Intelligence in Medicine

Background:

• Amyloid-beta (Aβ) plaque accumulation is a hallmark of Alzheimer's disease (AD).
• Accurate quantification of Aβ load is crucial for diagnosis and monitoring disease progression.
• Current methods using standardized uptake value ratio (SUVr) may have limitations in precision.

Purpose of the Study:

• To validate a novel method for quantifying Aβ load using nonspecific (NS) estimates from convolutional neural networks (CNNs).
• To assess the performance of this method using [18F]Florbetapir PET scans from longitudinal and multicenter Alzheimer's Disease Neuroimaging Initiative (ADNI) data.
• To compare the proposed method against the conventional SUVr for association with cognitive decline.

Main Methods:

• Utilized 188 paired MR (T1/T2-weighted) and PET images from the ADNI3 dataset.
• Trained multimodal ScaleNet (SN) and monomodal HighRes3DNet (HRN) to map structural MR to NS-PET images.
• Calculated specific amyloid load (SAβL) by subtracting estimated NS from SUVr images and evaluated its association with cognitive scores.

Main Results:

• SAβL derived from both SN and HRN demonstrated a stronger association with memory-related cognitive scores than SUVr.
• For longitudinal data, SAβL estimated from multimodal SN consistently outperformed SUVr across all memory-related cognitive tests.
• The CNN-based method successfully estimated NS for [18F]Florbetapir scans.

Conclusions:

• The proposed CNN-based method for Aβ load quantification correlates better with cognitive decline than SUVr, for both static and longitudinal data.
• Multimodal networks incorporating both T1-weighted and T2-weighted MR images are recommended for improved NS estimation.
• This approach offers a promising advancement in Aβ quantification for AD research and clinical applications.