Bayesian methods for FMRI time-series analysis using a nonstationary model for the noise.

Area of Science:

• Neuroimaging
• Statistical Analysis
• Signal Processing

Background:

• Functional MRI (fMRI) data analysis presents challenges due to non-stationary noise.
• Traditional methods may not fully capture the complex noise structure inherent in fMRI time series.

Purpose of the Study:

• To propose and evaluate novel Bayesian algorithms for fMRI data analysis.
• To address the issue of non-stationary noise in fMRI data using temporal and spatiotemporal approaches.
• To compare the performance of the proposed algorithms against existing methods like weighted least-squares (WLS).

Main Methods:

• Utilizing a Bayesian framework for fMRI data analysis.
• Developing two algorithms: one based on temporal analysis and another on spatiotemporal analysis.
• Employing variational Bayesian (VB) methodology to estimate regression parameters and noise variance components within a generalized linear model (GLM).
• Incorporating an extended design matrix in the GLM to account for drift in fMRI time series.

Main Results:

• The proposed VB-based algorithms effectively estimate noise variance across images and voxels.
• Iterative estimation of regression coefficients and variance components is achieved in a fully automated manner.
• Simulated and real fMRI data demonstrated the superiority of the proposed Bayesian approach over the WLS method.
• The algorithms successfully account for the complex, non-stationary noise structure in fMRI data.

Conclusions:

• The developed Bayesian algorithms offer a superior approach for fMRI data analysis compared to WLS.
• The VB methodology provides an efficient and automated way to handle complex noise in fMRI.
• These methods enhance the reliability and accuracy of fMRI data interpretation by addressing noise non-stationarity.