Modelling brain tissues intensities using Dirichlet process

Abstract

The work aimed to obtain a statistical model for predicting the intensities of the pixels for the different types of brain tissue in magnetic resonance images using Gaussian mixture models by Dirichlet process (GMMDP). The experimental work shows the implementation of the Gaussian mixture model by Dirichlet process to perform the segmentation of brain tissue. The result of the segmentation performed by Gaussian mixture models (GMM) by Dirichlet process for the brain tissues, compared to a parametric prediction method, indicates that targeting Gaussian mixture models by Dirichlet process achieves better results, from 2% to 7% for different tissues, although with limitations. Checking these limitations, we found that their cause was the overlap of the brain tissue intensities. In order to decrease these limitations we modified the model to fit the intensities of the brain tissue. In the case of cerebrospinal fluid tissue, we modified the model to work with Rice distribution. It was observed that with this change, the proposed segmentation improved but the difference between segmentations was not statistically significant. In the case of white matter tissue, we modified the model to work with new distributions: Genlogistic, Powernorm, Log Gamma, Gumbel skewed to the left and Johnson Su. It was observed that only the modified model that works with the Johnson Su distribution for the intensities of white matter is the one that can improve the segmentation of those brain tissues but the difference between segmentations was not statistically significant.