Alzheimer's disease classification using mutual information generated graph convolutional network for functional MRI.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• High-order cognitive functions rely on inter-regional brain communication, quantifiable by mutual information (MI).
• Alzheimer's disease (AD) impairs cognitive functions, suggesting alterations in inter-regional MI that require investigation.
• Current understanding of AD's impact on inter-regional brain communication, specifically MI, remains limited.

Purpose of the Study:

• To determine if inter-regional MI can differentiate Alzheimer's disease (AD) stages from normal controls (NC).
• To employ a connectome-based graph convolutional network (GCN) for predicting AD stages using MI.
• To compare the efficacy of MI-based connectomes against other connectivity measures.

Main Methods:

• Mutual information (MI) was computed between brain region time series to create connectomes.
• A multi-level connectome-based GCN (MLC-GCN) processed these connectomes for spatio-temporal feature extraction.
• The model was validated on 552 subjects from ADNI and OASIS3 datasets using 5-fold cross-validation.

Main Results:

• The MI-based connectome achieved superior prediction accuracy (87.72% ADNI2, 84.11% OASIS3) and Area Under the Curve (0.96 for both).
• Key MI features were identified in the temporal, prefrontal, and parietal cortices.
• MI-based models outperformed models using Kullback-Leibler divergence, cross-entropy, cross-sample entropy, and correlation coefficients.

Conclusions:

• MI-based connectomes reliably differentiate between normal controls, mild cognitive impairment, and Alzheimer's disease.
• Mutual information demonstrates superior performance compared to other connectivity measures for AD detection.
• Further validation on independent datasets is recommended for the MI-based connectome model.