Biomarkers.

Area of Science:

• Neuroimaging and Artificial Intelligence
• Biomarker Discovery for Neurodegenerative Diseases

Background:

• Alzheimer's disease (AD) diagnosis relies on detecting amyloid-β accumulation, often quantified using Positron Emission Tomography (PET) imaging.
• The high cost and limited availability of PET scans hinder widespread clinical application for AD assessment.
• Existing deep learning models for amyloid prediction often require complete datasets or actual PET images, limiting their real-world utility.

Purpose of the Study:

• To develop a masked multimodal-multitask deep learning framework for predicting amyloid burden in Alzheimer's disease.
• To integrate synthetic PET scans generated from MRI to overcome data limitations in routine clinical practice.
• To improve the accuracy and accessibility of early Alzheimer's disease screening.

Main Methods:

• A cohort of 968 participants from ADNI-2 and ADNI-3 studies (2,043 observations) with longitudinal MRI, PET, demographic, and clinical data was analyzed.
• A latent diffusion model (LDM) was employed to generate synthetic AV45-PET scans from MRI sequences.
• A deep learning network utilized synthetic PET images and available clinical data via a masked embedding attention mechanism to predict amyloid SUVRs and classify positivity, handling missing inputs.

Main Results:

• The proposed framework achieved a mean absolute error (MAE) of 0.11 for continuous SUVR prediction, outperforming baseline models (0.20, 0.13, 0.13).
• For amyloid positivity classification (SUVR > 1.11), the model achieved an area under the curve (AUC) of 0.93, surpassing baseline models (0.48, 0.89, 0.90).
• The approach effectively handled missing clinical data and reduced reliance on actual PET scans.

Conclusions:

• The developed deep learning framework significantly enhances the prediction of amyloid burden by integrating synthetic PET data and managing missing information.
• This approach has the potential to broaden access to early Alzheimer's disease screening in diverse clinical settings by reducing the dependency on costly PET imaging.
• Future research will focus on validation in larger, diverse cohorts and exploring applications with different imaging tracers for broader real-world impact on AD management.