Leveraging Brain Modularity Prior for Interpretable Representation Learning of fMRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for analyzing brain disorders by reflecting spontaneous neural activity.
• Current machine/deep learning methods for fMRI analysis often lack biological interpretability.
• The brain's modular structure in functional networks is not fully utilized by existing learning-based approaches.

Purpose of the Study:

• To propose a novel framework for interpretable fMRI analysis using brain modularity.
• To develop a dynamic representation learning method constrained by neurocognitive modules.
• To identify interpretable biomarkers for clinical diagnosis of brain disorders.

Main Methods:

• A brain modularity-constrained dynamic representation learning framework was developed.
• This framework includes dynamic graph construction and a modularity-constrained graph neural network (MGNN).
• The MGNN incorporates constraints from three core neurocognitive modules (salience, central executive, default mode networks) and a graph topology reconstruction constraint.

Main Results:

• The proposed method effectively learns interpretable representations from rs-fMRI data.
• Experimental results on 534 subjects across two datasets validate the framework's effectiveness.
• Discriminative brain regions-of-interest (ROIs) and functional connectivities were identified as potential biomarkers.

Conclusions:

• The brain modularity-constrained framework enhances the interpretability of fMRI analysis.
• The identified biomarkers show promise for aiding in the clinical diagnosis of brain disorders.
• This approach offers a more biologically informed way to analyze brain function using rs-fMRI.