Modularity-Constrained Dynamic Representation Learning for Interpretable Brain Disorder Analysis with Functional MRI.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Biomarker Discovery

Background:

• Resting-state functional MRI (rs-fMRI) is crucial for detecting brain disorders via functional connectivity.
• Current machine/deep learning methods for fMRI lack interpretability and fail to capture brain modularity.
• Understanding brain modularity is key to objective quantification of brain pathology.

Purpose of the Study:

• To propose a novel modularity-constrained dynamic representation learning (MDRL) framework for interpretable rs-fMRI analysis.
• To develop a method that effectively characterizes brain modularity and enhances biomarker interpretability.
• To improve the accuracy and explainability of brain disorder detection using rs-fMRI data.

Main Methods:

• Developed a modularity-constrained dynamic representation learning (MDRL) framework.
• Employed dynamic graph construction and a modularity-constrained spatiotemporal graph neural network (MSGNN).
• Integrated prediction and biomarker detection with MSGNN constrained by key functional modules (central executive, salience, default mode networks).

Main Results:

• The MDRL framework demonstrated superior performance compared to state-of-the-art methods across three datasets (1,155 subjects).
• The method effectively learned dynamic spatiotemporal representations of fMRI data.
• Detected fMRI biomarkers exhibited enhanced explainability, aiding in clinical diagnosis.

Conclusions:

• The proposed MDRL framework offers an interpretable approach for brain disorder analysis using rs-fMRI.
• This method has the potential to significantly improve objective quantification and clinical diagnosis of brain pathologies.
• The focus on brain modularity and dynamic feature learning represents a significant advancement in neuroimaging biomarker discovery.