MLC-GCN: Multi-Level Generated Connectome Based GCN for AD Detection.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Resting-state fMRI (rsfMRI) is crucial for Alzheimer's Disease (AD) research but faces challenges with feature extraction and noise.
• Existing graph convolution network (GCN) models struggle with insufficient feature representation and interpreting biological insights.

Purpose of the Study:

• To develop a novel multi-level connectome-generated GCN (MLC-GCN) for enhanced feature extraction in individual connectomes.
• To improve the accuracy and interpretability of AD detection using rsfMRI data.

Main Methods:

• Constructing multiple parallel connectomes using stacked spatiotemporal feature extractors (STFEs) to capture hierarchical features and reduce noise.
• Inputting each generated connectome into a GCN for advanced feature extraction.
• Concatenating GCN outputs for a multilayer perceptron to predict AD stages.

Main Results:

• MLC-GCN demonstrated superior performance in differentiating between normal controls, mild cognitive impairment, and AD.
• Validated on independent ADNI and OASIS-3 datasets, outperforming existing GCN architectures and AD classifiers.
• The model revealed high interpretability in identifying clinically relevant connectome nodes and connectivity features.

Conclusions:

• The proposed MLC-GCN effectively enhances feature extraction for individual connectomes, leading to improved AD detection.
• MLC-GCN offers a promising, interpretable approach for AD diagnosis and biomarker discovery using rsfMRI.
• This method advances the application of GCNs in neuroimaging for neurological disorder classification.