MML based energy minimization method for PVE classification in MR images

Partial volume effect (PVE) arises in medical images when more than one tissue type contributes to the pixel intensity. It is observed that pixel intensity of a PVE pixel depends on the proportions of each tissue type present. It is very important to estimate the partial volume effect, in order to increase the accuracy of the quantitative measurements made from magnetic resonance images (MRI). The main objective of this paper is to classify the PVE pixels present in a MR image with low probability of error by using information theoretic penalization to the log-likelihood of the given image and address the model initialization issue properly. We employ mixture models for this purpose, which are flexible and powerful probabilistic modeling tool for univariate and multivariate data. There usefulness is acknowledged in any area which involves statistical modeling of data, for example, image analysis, machine learning, computer vision and medical imaging. Furthermore, we use MML based penalization for unsupervised model selection and estimation of model parameters is accelerated by using noninformative initializations and component-wise EM. Validation is performed on a real MR image and its results are provided in this paper.