Spatio-Temporal Graph Convolution for Resting-State fMRI Analysis.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Resting-state fMRI (rs-fMRI) captures brain's intrinsic functional networks via Blood-Oxygen-Level-Dependent (BOLD) signals.
• Current deep learning models often overlook temporal dynamics or inter-regional dependencies in rs-fMRI data.

Purpose of the Study:

• To develop a deep learning approach that integrates both spatial and temporal information from rs-fMRI.
• To improve the accuracy of predicting demographic information (gender, age) from brain activity patterns.

Main Methods:

• Formulation of functional connectivity networks as spatio-temporal graphs.
• Training a spatio-temporal graph convolutional network (ST-GCN) on BOLD time series sub-sequences.
• Modeling non-stationary functional connectivity and learning edge importance for predictive insights.

Main Results:

• ST-GCN significantly outperformed existing methods in predicting gender and age using rs-fMRI data from HCP and NCANDA cohorts.
• Identified brain regions and functional connections critical for accurate predictions.
• Model-derived markers align with established neuroscience findings.

Conclusions:

• Spatio-temporal graph convolutional networks offer a powerful framework for analyzing complex rs-fMRI data.
• The ST-GCN approach enhances understanding of brain functional connectivity and its relation to demographic variables.
• This method holds potential for advancing neuroimaging-based biomarkers.