L0-regularized time-varying sparse inverse covariance estimation for tracking dynamic fMRI brain networks

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Network Science

Background:

• Understanding dynamic brain mechanisms requires exploring time-varying functional brain connectivity using functional Magnetic Resonance Imaging (fMRI).
• Inferring sparse functional brain networks commonly uses l1-penalized inverse covariance, with extensions for time-varying networks via sliding windows and temporal smoothing.
• l1 penalty has limitations in achieving maximum sparsity compared to l0 penalty, suggesting potential for improved inverse covariance estimation.

Purpose of the Study:

• To introduce a novel time-varying sparse inverse covariance estimation method utilizing dual l0-penalties (DLP).
• To enhance the estimation of sparse functional brain networks by leveraging the superior sparsity-inducing properties of the l0 penalty.

Main Methods:

• Developed a new DLP method that estimates sparse inverse covariance by minimizing an l0-penalized log-likelihood function.
• Incorporated an additional l0 penalty on temporal homogeneity to improve network structure estimation over time.
• Utilized a cyclic descent optimization algorithm to efficiently find the minimum of the objective function.

Main Results:

• The proposed DLP method demonstrated superior performance in estimating time-varying sparse network structures.
• Comparative experiments on simulated signals showed DLP outperforms conventional l1-penalized methods.
• DLP achieved better accuracy under various scenarios for dynamic functional brain connectivity.

Conclusions:

• The novel DLP method provides a more effective approach for estimating time-varying sparse brain networks.
• Dual l0-penalties offer advantages over l1-penalized methods for dynamic functional connectivity analysis.
• This method advances the understanding of dynamic brain mechanisms through improved network inference.