Predicting Brain Amyloid-β PET Grades with Graph Convolutional Networks Based on Functional MRI and Multi-Level...

Area of Science:

• Neuroimaging
• Alzheimer's Disease Research
• Artificial Intelligence in Medicine

Background:

• Amyloid-β (Aβ) deposition detection is key for Alzheimer's disease (AD) diagnosis.
• Current PET scans for Aβ have limitations in visual inspection and cost.

Purpose of the Study:

• To define non-binary Aβ deposition levels using PET data clustering.
• To assess the potential of non-invasive fMRI for predicting individual Aβ deposition grades.

Main Methods:

• Clustered Aβ-PET images (N=258) into three grades using t-SNE and k-means.
• Constructed functional connectivity (FC) networks from resting-state fMRI.
• Employed graph convolutional networks (GCNs) to predict Aβ-PET grades from FC network topology.

Main Results:

• Identified three distinct Aβ-PET grades with significant demographic and clinical differences.
• Achieved 78.8% accuracy in predicting Aβ-PET grades using GCNs on FC data.
• Demonstrated significant variations in gender, age, cognition, and APOE type across grades.

Conclusions:

• Functional MRI combined with deep learning shows promise for Aβ deposition grading.
• This non-invasive approach could complement or offer an alternative to costly PET scans.
• The findings support the feasibility of using fMRI for approximating PET-based Aβ grading in AD.