Improved EM-based tissue segmentation and partial volume effect quantification in multi-sequence brain MRI

Abstract

The Expectation Maximization algorithm is a powerful probabilistic tool for brain tissue segmentation. The framework is based on the Gaussian mixture model in MRI, and employs a probabilistic brain atlas as a prior to produce a segmentation of white matter, grey matter and cerebro-spinal fluid (CSF). However, several artifacts can alter the segmentation process. For example, CSF is not a well defined class because of the large quantity of voxels affected by the partial volume effect which alters segmentation results and volume computation. In this study, we show that ignoring vessel segmentation when handling partial volume effect can also lead to false results, more specifically to an over-estimation of the CSF variance in the intensity space. We also propose a more versatile method to improve tissue classification, without a requirement of any outlier class, so that brain tissues, especially the cerebro-spinal fluid, follows the Gaussian noise model in MRI correctly. © Springer-Verlag Berlin Heidelberg 2004.