MVS-GCN: A prior brain structure learning-guided multi-view graph convolution network for autism spectrum disorder...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Psychiatry

Background:

• Functional brain networks (FBN) show promise for classifying neurological disorders like Autism Spectrum Disorders (ASD).
• Diagnosing neurological disorders using FBN is challenging due to subject heterogeneity and network noise.
• Existing deep learning models often lack interpretability for brain network analysis.

Purpose of the Study:

• To develop an interpretable machine learning framework for neurological disorder classification using FBN.
• To address the challenges of heterogeneity and noise in brain network data.
• To improve the accuracy and interpretability of deep learning models in neurological disorder diagnosis.

Main Methods:

• A prior brain structure learning-guided multi-view graph convolutional neural network (MVS-GCN) was developed.
• The MVS-GCN model integrates graph structure learning and multi-task graph embedding learning.
• The approach aims to learn effective end-to-end representations for brain networks.

Main Results:

• The MVS-GCN method achieved enhanced performance on the Autism Brain Imaging Data Exchange (ABIDE) and Alzheimer's Disease Neuroimaging Initiative (ADNI) datasets.
• On the ABIDE dataset, MVS-GCN obtained an average accuracy of 69.38% and AUC of 69.01%.
• Results demonstrated consistency with prior neuroimaging evidence of ASD biomarkers and identified potential functional subnetworks.

Conclusions:

• The MVS-GCN method effectively addresses brain network heterogeneity and enhances feature representation for improved neurological disorder classification.
• The model captures essential embeddings for better diagnostic performance.
• The developed framework offers an interpretable approach to brain disorder diagnosis using functional brain networks.