Learning Brain Dynamics of Evolving Manifold Functional MRI Data Using Geometric-Attention Neural Network.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Brain network functional connectivities (FC) exhibit complex geometric patterns crucial for understanding brain dynamics.
• Characterizing time-evolving brain states requires advanced methods to analyze functional neuroimaging data.

Purpose of the Study:

• To present a novel geometric-attention neural network for characterizing time-evolving brain state changes.
• To track functional dynamics on high-dimensional Riemannian manifolds of symmetric positive definite (SPD) matrices.
• To learn state-specific manifold signatures representing underlying cognition.

Main Methods:

• Developed a neural network utilizing a Riemannian manifold of SPD matrices to model brain state evolution.
• Employed a Convolutional Neural Network (CNN) on the SPD manifold for low-dimension feature representation.
• Integrated an end-to-end Recurrent Neural Network (RNN) for time-varying mapping and a geometric attention mechanism within the CNN.

Main Results:

• The proposed manifold-based neural network successfully predicted brain state changes in both simulated and real neuroimaging data.
• The geometric attention mechanism identified latent geometric patterns in SPD matrices associated with brain states.
• Achieved promising results on task functional neuroimaging data from the Human Connectome Project.

Conclusions:

• The novel geometric-attention neural network effectively analyzes brain dynamics and predicts state changes.
• This approach offers a powerful tool for investigating the relationship between brain function, cognition, and behavior.
• Demonstrates significant applicability in advancing neuroscience research and understanding brain function.