A plug-in graph neural network to boost temporal sensitivity in fMRI analysis

TL;DR

This paper introduces GraphCorr, a plug-in graph neural network that enhances the temporal sensitivity of fMRI analysis by capturing dynamic brain activity features through transformer-based embedding and lag filtering.

Contribution

The proposed GraphCorr module can be integrated into existing fMRI models to improve their ability to detect dynamic brain activity patterns and interactions.

Findings

Improved classification accuracy on public fMRI datasets.

Enhanced interpretability of brain activity features.

Compatibility with multiple state-of-the-art models.

Abstract

Learning-based methods have recently enabled performance leaps in analysis of high-dimensional functional MRI (fMRI) time series. Deep learning models that receive as input functional connectivity (FC) features among brain regions have been commonly adopted in the literature. However, many models focus on temporally static FC features across a scan, reducing sensitivity to dynamic features of brain activity. Here, we describe a plug-in graph neural network that can be flexibly integrated into a main learning-based fMRI model to boost its temporal sensitivity. Receiving brain regions as nodes and blood-oxygen-level-dependent (BOLD) signals as node inputs, the proposed GraphCorr method leverages a node embedder module based on a transformer encoder to capture temporally-windowed latent representations of BOLD signals. GraphCorr also leverages a lag filter module to account for delayed interactions across nodes by computing cross-correlation of windowed BOLD signals across a range of time lags. Information captured by the two modules is fused via a message passing algorithm executed on the graph, and enhanced node features are then computed at the output. These enhanced features are used to drive a subsequent learning-based model to analyze fMRI time series with elevated sensitivity. Comprehensive demonstrations on two public datasets indicate improved classification performance and interpretability for several state-of-the-art graphical and convolutional methods that employ GraphCorr-derived feature representations of fMRI time series as their input.