Deep reinforcement learning guided graph neural networks for brain network analysis.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Modern neuroimaging constructs human brains as connectomes, essential for understanding function and disease.
• Graph neural networks (GNNs) show promise for brain network analysis via feature aggregation and global pooling.
• Existing GNN methods use a fixed architecture, neglecting individual brain network complexities.

Purpose of the Study:

• To propose a novel brain network representation framework, BN-GNN, to optimize GNN architectures for individual brain networks.
• To address the limitation of fixed GNN layers in analyzing diverse brain network structures.
• To enhance the performance of GNNs in brain network analysis tasks.

Main Methods:

• Developed BN-GNN, a framework employing deep reinforcement learning (DRL).
• Utilized DRL to automatically predict the optimal number of feature propagations (GNN layers) for each brain network.
• Applied the framework to eight brain network disease analysis tasks.

Main Results:

• BN-GNN successfully searches for optimal GNN architectures tailored to individual brain networks.
• The framework demonstrates improved performance compared to traditional GNNs across eight disease analysis tasks.
• Achieved enhanced upper bounds of performance in brain network analysis.

Conclusions:

• BN-GNN offers a flexible and adaptive approach to brain network representation learning.
• Optimizing GNN architecture through DRL significantly boosts performance in analyzing brain structure and disease.
• This method advances the application of GNNs in neuroscience and brain disease research.