An extensible hierarchical graph convolutional network for early Alzheimer's disease identification

Area of Science:

• Neuroimaging and Artificial Intelligence
• Computational Neuroscience
• Medical Image Analysis

Background:

• Early identification of Alzheimer's disease (AD) is crucial for effective management.
• Multi-modal magnetic resonance imaging (MRI) data offers rich information for AD diagnosis.
• Analyzing gray matter atrophy and connectivity abnormalities is key for understanding AD progression.

Purpose of the Study:

• To develop an advanced method for early Alzheimer's disease identification using multi-modal MRI.
• To comprehensively analyze gray matter atrophy and structural/functional connectivity abnormalities in AD.
• To integrate image features and non-image information for improved diagnostic accuracy.

Main Methods:

• Proposed an extensible hierarchical graph convolutional network (EH-GCN) for early AD identification.
• Utilized a multi-branch residual network (ResNet) to extract image features from multi-modal MRI.
• Employed a graph convolutional network (GCN) to analyze structural and functional connectivity between brain regions-of-interests (ROIs), incorporating an optimized spatial GCN for population-level analysis.

Main Results:

• The EH-GCN achieved high classification accuracies: 88.71% for AD vs. Normal Control (NC), 82.71% for AD vs. Mild Cognitive Impairment (MCI), and 79.68% for MCI vs. NC.
• Extracted connectivity features revealed that functional abnormalities precede gray matter atrophy and structural connection degradation in AD.
• The method demonstrated high computational efficiency and extensibility for incorporating additional data modalities.

Conclusions:

• The EH-GCN provides a comprehensive approach to analyzing brain changes in different stages of AD.
• This method aids in understanding the interplay of gray matter atrophy, white matter integrity, and functional connectivity.
• The findings support the potential for developing novel clinical biomarkers for early AD detection.