Markov Chain Monte Carlo Random Effects Modeling in Magnetic Resonance Image Processing Using the BRugs Interface to WinBUGS
Main Article Content
Abstract
A common feature of many magnetic resonance image (MRI) data processing methods is the voxel-by-voxel (a voxel is a volume element) manner in which the processing is performed. In general, however, MRI data are expected to exhibit some level of spatial correlation, rendering an independent-voxels treatment inefficient in its use of the data. Bayesian random effect models are expected to be more efficient owing to their information-borrowing behaviour.
To illustrate the Bayesian random effects approach, this paper outlines a Markov chain Monte Carlo (MCMC) analysis of a perfusion MRI dataset, implemented in R using the BRugs package. BRugs provides an interface to WinBUGS and its GeoBUGS add-on. WinBUGS is a widely used programme for performing MCMC analyses, with a focus on Bayesian random effect models. A simultaneous modeling of both voxels (restricted to a region of interest) and multiple subjects is demonstrated. Despite the low signal-to-noise ratio in the magnetic resonance signal intensity data, useful model signal intensity profiles are obtained. The merits of random effects modeling are discussed in comparison with the alternative approaches based on region-of-interest averaging and repeated independent voxels analysis.
This paper focuses on perfusion MRI for the purpose of illustration, the main proposition being that random effects modeling is expected to be beneficial in many other MRI applications in which the signal-to-noise ratio is a limiting factor.
To illustrate the Bayesian random effects approach, this paper outlines a Markov chain Monte Carlo (MCMC) analysis of a perfusion MRI dataset, implemented in R using the BRugs package. BRugs provides an interface to WinBUGS and its GeoBUGS add-on. WinBUGS is a widely used programme for performing MCMC analyses, with a focus on Bayesian random effect models. A simultaneous modeling of both voxels (restricted to a region of interest) and multiple subjects is demonstrated. Despite the low signal-to-noise ratio in the magnetic resonance signal intensity data, useful model signal intensity profiles are obtained. The merits of random effects modeling are discussed in comparison with the alternative approaches based on region-of-interest averaging and repeated independent voxels analysis.
This paper focuses on perfusion MRI for the purpose of illustration, the main proposition being that random effects modeling is expected to be beneficial in many other MRI applications in which the signal-to-noise ratio is a limiting factor.