Deep Learning-Based Prediction of PET Amyloid Status Using MRI

Area of Science:

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

Background:

• Identifying amyloid-beta (Aβ) positivity is crucial for Alzheimer's disease (AD) clinical trials and treatments.
• Current methods like PET scans or cerebrospinal fluid sampling are invasive or costly.
• Previous MRI-based deep learning models using only T1-weighted (T1w) images showed limited performance.

Purpose of the Study:

• To develop and evaluate a deep learning model for predicting Aβ positivity using multi-contrast MRI.
• To compare the performance of a model using T1w images alone versus one using T1w and T2-FLAIR images.

Main Methods:

• Retrospective analysis of multi-contrast MRI and PET data from three public datasets (ADNI, OASIS3, A4).
• Training two EfficientNet models: one with T1w images only, and another with T1w and T2-FLAIR images.
• Performance assessment using AUC, accuracy, sensitivity, and specificity on internal and external validation cohorts.

Main Results:

• The multi-contrast model (T1w + T2-FLAIR) significantly outperformed the T1w-only model in internal testing (AUC: 0.67 vs. 0.61, accuracy: 0.63 vs. 0.59).
• The highest performance (AUC: 0.71) was observed in patients with mild cognitive impairment (MCI).
• Consistent performance was demonstrated in the external validation set (AUC: 0.65, accuracy: 0.62).

Conclusions:

• Incorporating T2-FLAIR images alongside T1w images in a deep learning approach significantly enhances the prediction of PET-determined amyloid status from MRI.
• This multi-contrast MRI-based deep learning method offers a promising, non-invasive alternative for identifying amyloid positivity in Alzheimer's disease research.