Edge-Boosted Graph Learning

Edge-Boosted Graph Learning for Functional Brain Connectivity Analysis
📄 Published in IEEE ISBI 2025
📚 Paper: "EDGE-BOOSTED GRAPH LEARNING FOR FUNCTIONAL BRAIN CONNECTIVITY ANALYSIS"

Overview

This repository contains the code implementation for the IEEE ISBI 2025 paper that proposes a novel Graph Neural Network (GNN) framework incorporating edge-to-edge (eFC) and node-to-node relationships to better analyze functional brain connectivity. Unlike traditional GNNs that focus solely on node attributes, this method co-embeds edge functional connectivity to capture dynamic brain interactions more effectively.

Key Features

• ✅ Edge Time Series (eTS) calculation to extract co-fluctuation patterns between brain regions
• ✅ Construction of Edge Functional Connectivity (eFC) matrices
• ✅ Co-embedding of node and edge attributes using a unified GNN architecture
• ✅ Performance benchmarking on ADNI and PPMI fMRI datasets
• ✅ Significant improvement over CNN, GCN, CRGNN, and MGNN baselines

Installation

git clone https://github.com/YOUR_USERNAME/EdgeBoostedGNN.git
cd EdgeBoostedGNN
pip install -r requirements.txt

⚠️ Note: Use Python ≥ 3.8 and ensure PyTorch with CUDA is properly configured for GPU training.

Requirements

All dependencies are listed in conda_requirements.txt.
You can install them using:

pip install -r requirements.txt

Running the Model

Navigate to the scripts/ folder and launch training:

python utill_gpu.py

Datasets

We evaluate on the following datasets:

• ADNI Dataset
• PPMI Dataset

All fMRI scans were preprocessed using fMRIPrep and converted into node and edge attribute matrices.

Model Architecture

• Dual-layer GCN: node aggregation and edge aggregation
• Co-embedding layer fuses node and edge features
• Dropout, batch norm, and ReLU activations included

Performance Summary

Dataset	Accuracy	Precision	F1 Score
ADNI	0.8000 ± 0.0876	0.8437 ± 0.0614	0.7659 ± 0.1181
PPMI	0.7083 ± 0.0572	0.6821 ± 0.0860	0.6700 ± 0.0587

Our method consistently outperformed CNN, GCN, CRGNN, and MGNN baselines.

Acknowledgments

This work was conducted as part of the 2024 NSF REU program at UNC Greensboro and supported by NSF Grant CNS-2349369.

Citation

@inproceedings{yang2024edgeboosted,
  title={Edge-Boosted Graph Learning for Functional Brain Connectivity Analysis},
  author={Yang, David and Abdelmegeed, Mostafa and Modl, John and Kim, Minjeong},
  booktitle={IEEE ISBI},
  year={2024}
}