Dynamic multi-site graph convolutional network for autism spectrum disorder identification | JoVE Visualize

Area of Science:

Neuroscience
Machine Learning
Medical Imaging Analysis

Background:

Multi-site learning is crucial for autism spectrum disorder (ASD) identification due to neuroimaging data heterogeneity across different medical sites.
Existing multi-site graph convolutional networks (MSGCNs) often overlook inter-site correlations, potentially leading to suboptimal ASD identification.
Current functional magnetic resonance imaging (fMRI) feature extraction methods struggle with varying time-series lengths in multi-site datasets.

Purpose of the Study:

To propose a Dual Graph based Dynamic Multi-Site Graph Convolutional Network (DG-DMSGCN) for enhanced multi-site ASD identification.
To address limitations in feature extraction for fMRI data with diverse time-series lengths.
To improve ASD identification accuracy by effectively leveraging correlations between different medical sites.

Main Methods:

Introduced a sliding-window dual-graph convolutional network (SW-DGCN) for simultaneous temporal and spatial feature extraction from fMRI data, accommodating varying series lengths.
Developed a dynamic multi-site graph convolutional network (DMSGCN) to aggregate features from multiple sites, explicitly considering inter-site correlations.
Evaluated the DG-DMSGCN framework on the public ABIDE I dataset, comprising data from 17 medical sites.

Main Results:

The proposed DG-DMSGCN framework demonstrated superior performance compared to state-of-the-art methods in multi-site ASD identification.
Significant improvements in identification accuracy were achieved, highlighting the effectiveness of the dynamic multi-site aggregation approach.
The method successfully handled fMRI data heterogeneity and varying time-series lengths from different sites.

Conclusions:

The DG-DMSGCN offers a promising approach for robust and accurate ASD identification from multi-site fMRI data.
The framework's ability to capture inter-site correlations and handle data heterogeneity suggests potential clinical utility for practical ASD diagnosis.
The developed method advances the field of neuroimaging analysis for identifying neurodevelopmental disorders.