Brain multigraph prediction using topology-aware adversarial graph neural network.

Area of Science:

• Neuroimaging
• Graph Theory
• Artificial Intelligence

Background:

• Brain graphs (connectomes) from MRI are crucial for understanding brain changes in disorders.
• Existing methods struggle to synthesize multiple brain graphs from a single source and often overlook local graph topology.

Purpose of the Study:

• To develop a novel deep learning framework for jointly predicting multiple brain graphs from a single source graph.
• To address limitations in scalability and topological preservation in current graph generation techniques.

Main Methods:

• Introduced a topology-aware graph Generative Adversarial Network (GAN) architecture (topoGAN).
• Employed a graph adversarial auto-encoder for multi-graph prediction and incorporated graph clustering to mitigate mode collapse.
• Integrated a topological loss function to ensure the prediction of topologically sound target brain graphs.

Main Results:

• topoGAN successfully predicts multiple brain graphs from a single source graph.
• The method demonstrates superior performance compared to baseline approaches across five target domains.
• Preservation of both global and local topological structures in the predicted brain graphs was achieved.

Conclusions:

• The proposed topology-aware GAN offers a scalable and effective solution for brain multigraph prediction.
• This advancement aids in characterizing brain disorders by providing more comprehensive connectome data.
• topoGAN sets a new standard for generating realistic and topologically accurate brain graph networks.