Residual graph transformer for autism spectrum disorder prediction

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on identifying reliable biomarkers.
• Resting-state functional magnetic resonance imaging (rs-fMRI) and brain functional connectivity (FC) show promise for ASD prediction.
• Existing methods face challenges in capturing complex brain interactions, deep network representation, and diagnostic interpretability.

Purpose of the Study:

• To propose a novel deep learning model, the FC-learned Residual Graph Transformer Network (RGTNet), for improved ASD prediction.
• To address limitations in harnessing brain region interactions, representation learning, and diagnostic interpretability.
• To develop a method for identifying clinically relevant biomarkers for ASD.

Main Methods:

• Utilizing rs-fMRI data to compute brain functional connectivity (FC) matrices.
• Designing a Graph Encoder to capture temporal dependencies and model interpretable FC matrices.
• Implementing a residual trick to deepen Graph Convolutional Network (GCN) architecture for higher-level feature learning.
• Employing Graph Sparse Fitting and weighted aggregation to manage dimensionality.

Main Results:

• RGTNet demonstrated superior performance on two ABIDE datasets compared to existing methods.
• Achieved an accuracy of 73.4% on the AAL atlas using five-fold cross-validation, outperforming the 70.9% benchmark.
• Identified biomarkers that align with established medical knowledge, supporting clinical relevance.

Conclusions:

• RGTNet offers a promising approach for enhancing ASD prediction through advanced functional connectivity analysis.
• The model's interpretability and biomarker identification capabilities pave the way for more reliable clinical diagnosis of ASD.
• The developed methodology provides a viable tool for advancing neuroimaging-based diagnostics in psychiatry.