Diffusion-based spatial priors for functional magnetic resonance images.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Modeling

Background:

• Conventional functional magnetic resonance imaging (fMRI) analysis in SPM involves pre-smoothing data, assuming uniform brain smoothness.
• This pre-smoothing prevents verification of the smoothness assumption against the actual data.
• Explicit spatial priors offer a data-driven approach to estimate smoothness.

Purpose of the Study:

• To apply a Bayesian scheme with diffusion-based spatial priors to fMRI data analysis.
• To investigate functional activations within the auditory system using a single-subject design.
• To demonstrate the computational efficiency and generalizability of diffusion-based priors.

Main Methods:

• Formulating spatial priors based on diffusion in terms of graph Laplacian eigenmodes.
• Utilizing eigenmodes with small eigenvalues for computational efficiency.
• Generalizing diffusion-based priors to encompass conventional Laplacian priors and Gaussian process models.
• Employing restricted maximum likelihood for covariance component estimation.

Main Results:

• Demonstrated that diffusion-based priors can be efficiently computed by discarding eigenmodes with small eigenvalues.
• Showcased diffusion-based priors as a generalization of Laplacian priors.
• Established diffusion-based priors as a special case of Gaussian process models.
• Provided strong evidence for a non-stationary spatial process in auditory fMRI data, contrasting with the stationary assumption of conventional smoothing.

Conclusions:

• Diffusion-based spatial priors offer a flexible and computationally efficient alternative to conventional smoothing in fMRI analysis.
• This method allows for data-driven estimation of spatial smoothness and formal model comparison.
• The findings support the existence of non-stationary spatial processes in the human auditory system.