Dynamic graph transformer network via dual-view connectivity for autism spectrum disorder identification

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Autism Spectrum Disorder (ASD) diagnosis requires objective methods for early intervention.
• Existing functional connectivity (FC) analysis methods overlook complementary information in positive and negative FCs and require pre-defined graphs.
• Challenges in ASD identification include class imbalance and multi-site data heterogeneity.

Purpose of the Study:

• To propose a novel dynamic graph Transformer network for accurate Autism Spectrum Disorder (ASD) identification.
• To leverage dual-view connectivity (positive and negative FCs) for enhanced information extraction.
• To address class imbalance and improve model generalizability in multi-site datasets.

Main Methods:

• Developed a dual-view feature extractor to capture individual and complementary information from positive and negative connectivity.
• Innovated a Graph Transformer network with a dynamic KNN module for graph construction without additional information.
• Implemented PolyLoss and Vrex methods to mitigate class imbalance and enhance generalizability.

Main Results:

• The proposed method achieved superior performance in ASD identification compared to state-of-the-art techniques.
• Demonstrated satisfying generalizability on the ABIDE I dataset with 1102 subjects.
• Effectively utilized both positive and negative functional connectivity for improved diagnostic accuracy.

Conclusions:

• The novel dynamic graph Transformer network offers a promising approach for objective and accurate ASD identification.
• Dual-view connectivity analysis significantly enhances the model's ability to differentiate between ASD and control subjects.
• The method shows potential for robust application in real-world clinical settings, addressing key limitations of previous approaches.