Bayesian co-segmentation of multiple MR images

Segmentation is one of the basic problems in magnetic resonance (MR) image analysis. We consider the problem of simultaneously segmenting multiple MR images, which, for example, can be a series of 2D/3D images of the same tissue scanned over time, different slices of a volume image, or images of symmetric parts. These multiple MR images share common structure information and hence they can assist each other in the segmentation procedure. We propose a Bayesian co-segmentation algorithm where the shared information across multiple images is utilized via a Markov random field prior. An efficient algorithm based on the Swendsen–Wang method is employed for posterior sampling, which is more efficient than the single-site Gibbs sampler. Because our co-segmentation algorithm pulls all the image information into consideration, it provides more accurate and robust results than individual segmentation, as supported by our experimental studies with real examples.

Keywords

MRI, co-segmentation, Bayesian, MCMC

Full Text (PDF format)

Published 1 January 2010