Functional Connectivity Networks with Latent Distributions for Mild Cognitive Impairment Identification

Area of Science:

• Neuroimaging
• Machine Learning
• Statistical Modeling

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) signals are complex and variable.
• Estimating brain functional connectivity networks is crucial for understanding brain function and dysfunction.
• Existing methods may struggle with the inherent variability of rs-fMRI data.

Purpose of the Study:

• To develop a novel generative learning approach for estimating brain functional connectivity networks.
• To model rs-fMRI signals as probability distributions for more robust connectivity estimation.
• To improve the identification of Mild Cognitive Impairment (MCI) using enhanced functional network analysis.

Main Methods:

• Utilized variational autoencoder (VAE) networks to encode rs-fMRI signals into confidence distributions in latent space.
• Mapped mean time series of brain regions of interest to multivariate Gaussian distributions.
• Measured pairwise brain region correlation using Jensen-Shannon divergence to create adjacency matrices representing functional connectivity.

Main Results:

• The proposed VAE-based method effectively estimates functional connectivity networks.
• Adjacency matrices derived from VAE latent spaces demonstrated complementarity for MCI identification.
• Classification performance for MCI detection was significantly improved by cascading classifiers.

Conclusions:

• The novel framework constructs brain functional networks from a statistical modeling perspective.
• The approach enhances statistical power for population data and generalization for data variability.
• Experimental results on a public dataset show superior performance compared to baseline and state-of-the-art methods for MCI identification.