fMRI Signal Prediction Using Spatio-Temporal Graph Neural Networks

This repository contains code for predicting fMRI signals using Spatio-Temporal Graph Neural Networks written in Julia.

The aim is to predict brain activity from fMRI data by leveraging Spatio-Temporal GNNs to model both spatial and temporal dependencies. This simple project was developed independently and later found to have similarities with a previously published paper titled "Forecasting brain activity based on models of spatiotemporal brain dynamics: A comparison of graph neural network architectures", validating the approach and exploring its potential.

Installation

1. Clone the repository
2. Install the required packages by running the following command in the Julia REPL:

using Pkg
Pkg.activate(".")
Pkg.instantiate()

Usage

The main script is training.jl which contains the code for training and evaluating the model. The script can be run using the following command in the Julia REPL:

include("src/training.jl")

Data

The data used in this project is from the Temporal Brain Networks dataset which contains a collection of temporal brain networks. The networks are obtained from resting-state fMRI data of 1047 subjects from the Human Connectome Project (HCP). The data is stored in the data/resting-state directory.

Model

The model used in this project is the Spatio-Temporal Graph Neural Network TGCN from the paper "T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction". The graph is constructed starting from the Schaefer 100 parcellation of the brain, and the fMRI signals are used as node features.

Results