Brain Magnetic Resonance Image Segmentation via Mixture Model Averaging

• Activity Number: 467
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Yichen Qin*+ and Kathryn I. Alpert and Carey Priebe and Lei Wang
• Companies: The Johns Hopkins University and Northwestern University and The Johns Hopkins University and Northwestern University
• Keywords: Gaussian mixture model ; Magnetic resonance images segmentation ; Bayesian information criterion ; Model averaging ; Model selection

Abstract

In this article, we show an application of the mixture model for the segmentation of brain Magnetic Resonance Images (MRI). The goal is to segment the brain MR images into three parts --- cerebrospinal fluid, gray matter and white matter, which will bring us more evidence on the association between the abnormality of the brain structure and Alzheimer disease. This application is built on the methodology discussed in Priebe et al. 2006 and Lee et al. 2008, but with modifications on the design of the classifier. Instead of estimating the mixture model complexity, we propose an ensemble learning approach --- model averaging --- so as to get more robust estimation. We first fit multiple mixture models to the data with model complexity specified ex ante, and then use Bayesian Information Criterion (BIC) to weight average all the models to get more robust segmentation results. The proposed methodology is well explained and is also validated in the application of segmenting real brain cingulate gyrus MRI data. We also show the comparison between the model averaging results and single model results in terms of robustness.