# Deep Reinforcement Learning Guided Graph Neural Networks for Brain Network Analysis

**Authors:** Xusheng Zhao, Jia Wu, Hao Peng, Amin Beheshti, Jessica Monaghan, David McAlpine, Heivet Hernandez-Perez, Mark Dras, Qiong Dai, Yangyang Li, Philip S. Yu, Lifang He  
**arXiv:** [2203.10093](https://arxiv.org/abs/2203.10093)  
**Date:** 2022-03-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.10093v1/

## Key points

- The study focuses on applying modern neuroimaging techniques like DTI and fMRI to model the human brain as a connectome.
- Understanding structural information and hierarchical patterns of brain networks is crucial for deciphering brain functions and disease states.
- Graph neural networks (GNNs) are increasingly used for brain network analysis due to their promising capabilities in network representation learning.
- Existing GNN-based methods may overlook the need for varying aggregation iterations for different subjects' brain networks and typically use a fixed number of layers, limiting their potential.
- Researchers propose a novel framework called BN-GNN that leverages deep reinforcement learning to determine optimal feature aggregations for individual brain network instances, aiming to enhance traditional GNN performance.
- BN-GNN outperforms traditional GNNs in various brain network analysis tasks based on experiments conducted on eight real-world datasets, showing its potential to improve understanding of complex brain functions and disease mechanisms.

## Summary

The study focuses on the application of modern neuroimaging techniques such as diffusion tensor imaging (DTI) and functional magnetic resonance imaging (fMRI) to model the human brain as a brain network or connectome. Understanding the structural information and hierarchical patterns of brain networks is crucial for deciphering brain functions and disease states. Recently, there has been a surge in interest in utilizing graph neural networks (GNNs) for brain network analysis due to their promising network representation learning capabilities. However, existing GNN-based methods often overlook the fact that different subjects may require varying aggregation iterations for their brain networks. Additionally, these methods typically use GNN with a fixed number of layers to learn all brain networks, which may not fully exploit the potential of GNNs for promoting comprehensive brain network analysis. To address these limitations, the researchers propose a novel brain network representation framework called BN-GNN. This framework leverages deep reinforcement learning (DRL) to train a meta-policy that automatically determines the optimal number of feature aggregations (corresponding to the number of GNN layers) required for each specific brain network instance. By customizing the GNN architecture for individual brain networks, BN-GNN aims to enhance the performance of traditional GNNs on various brain network analysis tasks. The study conducts extensive experiments on eight real-world brain network datasets to evaluate the effectiveness of BN-GNN. The results demonstrate that BN-GNN outperforms traditional GNNs in different brain network analysis tasks, showcasing its potential to improve our understanding of complex brain functions and disease mechanisms. In conclusion, this research introduces an innovative approach that tailors GNN architectures to individual brain networks through deep reinforcement learning, offering a promising avenue for advancing neuroimaging research and enhancing our insights into the intricate workings of the human <kw>brain</kw>.