Combined spatial and non-spatial prior for inference on MRI time-series

Area of Science:

• Neuroimaging
• Statistical Modeling
• Biophysics

Background:

• Generalized Linear Models (GLM) are standard for fMRI/MRI time-series, but lack flexibility.
• Bayesian methods offer principled prior incorporation, yet combining spatial and fixed priors is challenging.

Purpose of the Study:

• Develop a Gaussian-process-based prior for adaptive spatial regularization in MRI analysis.
• Enable principled integration of both spatial and fixed biophysical priors on parameters.
• Improve parameter estimation in fMRI and ASL data analysis.

Main Methods:

• Utilized a Gaussian-process-based prior with a parameterized covariance matrix for adaptive spatial regularization.
• Employed evidence optimization (EO) and variational Bayes (VB) for parameter updates.
• Applied a linear Taylor expansion for non-linear forward models.

Main Results:

• Demonstrated improved fMRI model fits with a constrained hemodynamic response function (HRF) model compared to separate priors.
• Successfully analyzed multi-inversion arterial spin labeling (ASL) data using a non-linear perfusion model.
• Showcased consistent performance and superior estimates when combining spatial and biophysical priors.

Conclusions:

• The developed Gaussian-process prior effectively integrates spatial and biophysical information in MRI time-series analysis.
• This unified approach enhances model accuracy and provides more reliable parameter estimates across diverse neuroimaging applications.
• Offers a robust alternative to standard GLM, particularly when complex prior information is available.