Siamese Graph Convolutional Network quantifies increasing structure-function discrepancy over the cognitive decline continuum | JoVE Visualize

Area of Science:

Neuroscience
Machine Learning
Medical Imaging

Background:

Alzheimer's disease dementia (ADD) alters brain connectivity, but current measures lack diagnostic specificity.
Graph Neural Networks (GNNs) show promise in brain research, yet often use single-modality networks.
The interdependence of structural and functional brain connectivity is crucial but often overlooked in ADD research.

Purpose of the Study:

To develop a novel deep learning model, the structure-function discrepancy learning network (sfDLN), to quantify the disruption in brain connectivity in Alzheimer's disease.
To utilize the structure-function discrepancy score as a diagnostic biomarker for cognitive decline.
To compare the diagnostic performance of sfDLN against existing state-of-the-art classifiers.

Main Methods:

A siamese GNN architecture (sfDLN) was employed, hypothesizing increased structure-function mismatch with cognitive decline (Subjective Cognitive Impairment [SCI] to Mild Cognitive Impairment [MCI] to ADD).
Input data included structural brain connectomes (sNET) from diffusion MRI and sparse functional brain connectomes (ℓNET) from fMRI.
The sfDLN was trained to extract connectome representations and a discrepancy score, then blindly tested on an MCI cohort.

Main Results:

sfDLN generated discrepancy scores showed significant differences between ADD and SCI subjects.
Leave-one-out classification of SCI vs. ADD achieved 88% accuracy, outperforming existing GNN-based methods.
Blind assessment on MCI subjects confirmed sfDLN's ability to capture intermediate cognitive decline stages.
GNNExplainer analysis identified neurologically relevant cortical regions associated with ADD in the discrepancy scores.

Conclusions:

Brain structure-function harmony degrades with increasing cognitive decline, supporting the study's hypothesis.
The quantified discrepancy, rooted in ADD-relevant brain regions, serves as a potent diagnostic biomarker.
sfDLN demonstrates superior performance in classifying ADD compared to current GNN-based approaches.