A Mixture-Model Approach to Segmenting Magnetic Resonance Images

• Activity Number: 118
• Type: Topic Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: IMS
• Author(s): Ranjan Maitra*+
• Companies: Iowa State University
• Keywords: magnitude Magnetic Resonance Images ; initialization ; Kolmogorov test ; Rayleigh distribution

Abstract

Magnetic resonance (MR) images are usually magnitudes of complex-valued intensities at a voxel. We develop a practical model-based approach to segmenting three-dimensional MR images using a mixture of Riceans rather than a mixture of Gaussians that has traditionally been assumed. Specifically, we develop a version of the expectation-maximization (EM) algorithm by introducing the discarded phase information at each voxel as missing observations. The complete likelihood is then a member of the regular exponential family, and the EM can then be implemented without the need for potentially unstable numerical optimization methods. Spatial context to the segmentation is also introduced in our mixture model. An added benefit of our approach above is the ready estimation of the variability in the estimate, which can potentially be used to provide quantification in our diagnosis. We evaluate our methodology on both realistic simulation datasets as well as images acquired on a physical phantom and on human subject.