Bayesian Rician Regression for Neuroimaging.

Area of Science:

• Medical Imaging
• Statistical Modeling
• Neuroscience

Background:

• Diffusion Weighted Imaging (DWI) data typically follows a Rician distribution.
• This distribution is also pertinent for functional Magnetic Resonance Imaging (fMRI) data acquired at high resolutions.
• Existing models often approximate Rician noise with a Gaussian model, potentially leading to inaccuracies.

Purpose of the Study:

• To propose a general regression model for non-central chi (NC-χ) distributed data.
• To introduce a heteroscedastic Rician regression model as a key application.
• To enable Bayesian variable selection for linking Rician distribution parameters to explanatory variables.

Main Methods:

• Development of a general regression model for NC-χ distributed data.
• Implementation of Bayesian variable selection to identify relevant explanatory variables.
• Utilization of a Markov Chain Monte Carlo (MCMC) algorithm for efficient posterior simulation and model uncertainty assessment.

Main Results:

• Simulated data demonstrated superior signal detection accuracy of the Rician model compared to the Gaussian model at low signal-to-noise ratios.
• Analysis of Human Connectome Project diffusion data revealed that the Gaussian noise model underestimates Mean Diffusivity (MD) and Fractional Anisotropy (FA).

Conclusions:

• The proposed Rician regression model offers a more accurate approach for analyzing diffusion imaging data than traditional Gaussian models.
• This improved modeling enhances the precision of diffusion parameter estimation, particularly in low signal-to-noise conditions.
• The findings have implications for more accurate quantification in diffusion MRI studies.