An Invertible Dynamic Graph Convolutional Network for Multi-Center ASD Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Autism Spectrum Disorder (ASD) diagnosis is challenging due to symptom variability.
• Current deep learning models for ASD classification using brain connectivity struggle with multi-center data, limited feature representation, and poor interpretability.

Purpose of the Study:

• To propose an interpretable deep learning model for Autism Spectrum Disorder (ASD) identification.
• To investigate brain connectivity alterations associated with ASD.
• To improve classification performance on multi-center data.

Main Methods:

• Developed an invertible dynamic Graph Convolutional Network (GCN) model.
• Incorporated invertible blocks for feature reconstruction and interpretability.
• Utilized pre-screening of connectivity features and a fully-connected layer for classification.

Main Results:

• Achieved superior disease classification performance on a dataset of 867 subjects.
• Demonstrated the model's ability to reconstruct input dynamic features, enhancing interpretability.
• Showcased improved performance compared to existing deep learning approaches for ASD.

Conclusions:

• The proposed invertible dynamic GCN offers an interpretable deep learning framework for brain connectivity analysis.
• This method holds significant potential for studying brain-related disorders like ASD.
• The approach enhances diagnostic accuracy and provides insights into disease-specific connectivity patterns.