Identification of autism spectrum disorder using multiple functional connectivity-based graph convolutional network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) combined with graph convolutional networks (GCN) shows promise for early autism spectrum disorder (ASD) diagnosis.
• Current GCN approaches often overlook crucial prior information from ASD-associated brain subnetworks, focusing solely on full-brain connectivity.

Purpose of the Study:

• To propose a novel Multiple Functional Connectivity-based Graph Convolutional Network (MFC-GCN) framework for improved ASD diagnosis.
• To incorporate both full-brain and key ASD-related brain subnetwork functional connectivity data into the GCN model.
• To address the heterogeneity within the Autism Brain Imaging Data Exchange (ABIDE) dataset using a novel External Attention Network Readout (EANReadout).

Main Methods:

• Development of the MFC-GCN framework integrating full-brain and subnetwork functional connectivity.
• Introduction of the EANReadout mechanism to handle dataset heterogeneity and explore subject associations.
• Experimental validation on the ABIDE dataset comprising 714 subjects.

Main Results:

• The proposed MFC-GCN framework achieved an average accuracy of 70.31% on the ABIDE dataset.
• The novel EANReadout significantly outperformed traditional readout layers.
• The EANReadout improved the overall framework accuracy by 4.32%.

Conclusions:

• The MFC-GCN framework with EANReadout offers a more effective approach for early ASD diagnosis by leveraging multi-level brain connectivity information.
• The EANReadout is crucial for managing data heterogeneity and enhancing classification performance in ASD research.
• This study highlights the potential of advanced GCN techniques for neuroimaging-based disorder diagnosis.