A Hierarchical Graph Convolutional Network With Infomax-Guided Graph Embedding for Population-Based ASD Detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Functional magnetic resonance imaging (fMRI)-based brain networks show promise for detecting autism spectrum disorders (ASD).
• Graph convolution networks (GCNs) have improved ASD classification using fMRI data.
• Existing GCN methods often underutilize brain functional connectivity network (BFCN) topology and non-imaging data.

Purpose of the Study:

• To develop a novel hierarchical graph embedding model for improved ASD detection.
• To integrate both BFCN topological information and subject non-imaging data.
• To enhance the accuracy and robustness of ASD diagnostic tools.

Main Methods:

• Proposed a hierarchical graph embedding model incorporating an Infomax Module for feature extraction from brain regions of interest (ROIs).
• Constructed a population graph model using extracted fMRI features and non-imaging data.
• Employed a graph convolution framework for feature propagation and aggregation for ASD detection.
• Utilized the Autism Brain Imaging Data Exchange (ABIDE) dataset for model evaluation.

Main Results:

• Achieved an average accuracy of 77.2% for autism detection.
• Obtained an Area Under the Curve (AUC) of 87.2%.
• Demonstrated superior performance compared to baseline approaches.

Conclusions:

• The proposed model effectively integrates BFCN topology and non-imaging information for enhanced ASD detection.
• The model shows competitiveness, robustness, and effectiveness in aiding ASD diagnosis.
• This approach offers a promising advancement in the application of neuroimaging and AI for psychiatric disorders.