Scalable geometric learning with correlation-based functional brain networks.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Functional brain networks are crucial in neuroimaging, typically analyzed using correlation matrices.
• Existing methods often ignore the geometric properties of correlation structures, leading to inefficiencies.
• Previous geometric approaches face computational and stability issues, particularly with high-dimensional data.

Purpose of the Study:

• To develop a novel geometric framework for analyzing functional brain networks.
• To enable scalable, geometry-aware analyses of correlation matrices.
• To improve computational efficiency and numerical stability in high-dimensional neuroimaging data.

Main Methods:

• Proposed a novel geometric framework using diffeomorphic transformations.
• Embedded correlation matrices into Euclidean space while preserving manifold characteristics.
• Integrated the framework with standard machine learning techniques (regression, dimensionality reduction, clustering).

Main Results:

• Demonstrated significant improvements in computational speed and predictive accuracy via simulations.
• Showcased enhanced performance in behavioral score prediction, subject fingerprinting (resting-state fMRI), and EEG hypothesis testing.
• Validated the framework's versatility and scalability on real neuroimaging data.

Conclusions:

• The proposed framework offers an efficient and interpretable geometric modeling approach for large-scale functional brain networks.
• This method enhances the integration of neuroimaging data with machine learning.
• Open-source tools are provided to facilitate community adoption and reproducibility.