Graph_convolutional_neural_network

Requirements

To run the code, please install the following dependencies:

pip install networkx
pip install torch-geometric

Torch Geometric is an extension library for PyTorch designed specifically for building Graph Neural Networks (GNNs). It follows the same class structure and data handling conventions as PyTorch, making it easy to integrate with existing PyTorch workflows.

Project Overview

This project aims to develop a potential biomarker for Autism Spectrum Disorder (ASD) using fMRI data and brain network configurations derived via BOLD signals. We construct whole-brain causal networks where:

• Nodes represent salient brain regions (defined by the EZ (Eickhoff-Zilles) atlas).
• Edges represent directed causal influences between regions.

We apply Graph Neural Networks (GNNs) to classify these brain networks. GNNs are designed to operate on data with an inherent graph structure, such as:

• fMRI signals aligned with anatomical atlases,
• social networks,
• and other non-Euclidean domains.

Unlike traditional CNNs, which rely on fixed grid-based kernels, GNNs aggregate features from neighboring nodes based on graph topology. This enables more flexible and biologically meaningful modeling of brain connectivity.

Methodology

Graph Construction

• Nodes: Brain regions based on the EZ atlas
• Edges: Top K nearest neighbors from the Pearson correlation matrix
• Node Features: fMRI neural activation signals

Models Used

• Graph Convolutional Network (GCN) Classifier
• Global Attention Network

These models were trained and evaluated using the ABIDE dataset.

References

• fMRI-based networks defines the performance of a graph neural network for the classification
• Benchmarking Graph Neural Networks for FMRI analysis