Collaborative learning of graph generation, clustering and classification for brain networks diagnosis

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Accurate diagnosis of autism spectrum disorder (ASD) is crucial for patient care.
• Functional brain networks (FBNs) offer insights into ASD, but high dimensionality and limited data pose challenges for classification models like graph convolutional networks (GCNs).

Purpose of the Study:

• To develop a robust framework for ASD diagnosis using FBNs.
• To address the limitations of high-dimensional connectivity and small sample sizes in brain network modeling.
• To enhance the classification performance of GCNs for ASD detection.

Main Methods:

• Introduced a unified three-stage graph learning framework: Graph Generation, Clustering, and Classification Networks (GraphCGC-Net).
• Employed multi-graph clustering (MGC) with a supervision scheme to enhance critical connections.
• Generated realistic brain networks by preserving global distribution and local topology.

Main Results:

• The GraphCGC-Net achieved 70.45% accuracy and 72.76% AUC on the Autism Brain Imaging Data Exchange (ABIDE) dataset.
• Demonstrated significant improvements over traditional GCN models, with a 9.3% increase in accuracy and 10.64% increase in AUC.
• MGC identified biologically meaningful subnetworks consistent with known ASD biomarkers.

Conclusions:

• GraphCGC-Net is effective for graph classification in diagnosing brain disorders like ASD.
• The framework's ability to generate biologically relevant subnetworks highlights its potential for biomarker discovery.
• Further investigation into using generative adversarial networks (GANs) for brain network classification is warranted.