SpaRG: Sparsely Reconstructed Graphs for Generalizable fMRI Analysis

Welcome to the repository for SpaRG, a cutting-edge framework that utilizes sparsely reconstructed graphs to enhance fMRI analysis by optimizing for sparsification and reconstruction. This repository contains the code, data, and results for our work on improving the interpretability and robustness of functional connectome analyses.

📚 Background

Resting-state fMRIs (rs-fMRIs) offer rich insights into psychiatric disorders and personal traits. However, analyzing functional connectomes—measured via BOLD signal correlations—is challenging due to their high dimensionality and susceptibility to variations.

🧠 SpaRG addresses these challenges by leveraging Graph Neural Networks (GNNs) and Variational Autoencoders (VAE) to sparsify inputs during training, ultimately improving the classification accuracy while dramatically reducing the number of neural connections considered.

🔬 Key Features

• Sparsification: Learn a sparse mask during training that identifies robust neural connections while discarding irrelevant ones.
• Reconstruction: Use a Variational Autoencoder (VAE) to reconstruct the connectome from the sparse input, allowing better interpretability.
• Classification: Feed the reconstructed connectomes into a GNN for the final classification task.
• Generalizability: Achieves strong results across both in-distribution (ID) and out-of-distribution (OOD) data sites.

📊 Results

SpaRG achieves up to 99% reduction in neural connections while maintaining or improving classification accuracy across both ID and OOD datasets.

	DiFuMo 64x64	DiFuMo 1024x1024
GCN	76.17±2.2	77.24±2.7
FCN	73.94±4.2	78.34±2.7
xGW-GAT	46.89±8.2	40.12±3.5
Mask-GCN	76.14±2.6	76.83±3.4
LASSO	82.10±8.4	83.74±2.4
ElasticNet	76.97±2.8	83.55±2.1
Frobenius	74.24±5.9	82.55±5.0
SpaRG (ours)	82.40±4.5	84.28±5.5

🛠️ Installation

To get started with SpaRG, follow these simple steps:

1. Clone the repository

2. Navigate to the project directory:

   cd SpaRG

3. Install the required dependencies:

   pip install -r requirements.txt

4. Download the ABIDE dataset (or your chosen dataset).

🚀 Usage

To train the model and perform classification, use the following command (along with other chosen arguments):

python3 main.py --model_name sparg --sparse_method vae

🧑‍🔬 Contributors

• Camila González (Stanford University)
• Yanis Miraoui (Stanford University)
• Yiran Fan (Stanford University)
• Ehsan Adeli (Stanford University)
• Kilian M. Pohl (Stanford University)

This work was partly funded by the U.S. National Institute of Health (NIH), the Stanford HAI Google Cloud Credit, and the DGIST Joint Research Project.

📜 Citation

If you use this code, please cite our paper:

@inproceedings{2024sparg,
  title={Sparsely Reconstructed Graphs for Generalizable fMRI Analysis},
  author={Camila González and Yanis Miraoui and Yiran Fan and Ehsan Adeli and Kilian M. Pohl},
  booktitle={MLCN 2024},
  year={2024}
}

📬 Contact

For any questions or collaboration opportunities, please reach out to us.