Electronic Theses and Dissertations





Document Type


Degree Name

Doctor of Philosophy


Electrical and Computer Engr


Electrical Engineering

Committee Chair

Iftekharuddin Khan

Committee Member

Olesegun Ebenezer George

Committee Member

David Russomanno

Committee Member

Robert Ogg


Posterior Fossa (PF) is a type of brain tumor located in or near brain stem and cerebellum. About 55% - 70 % pediatric brain tumors arise in the posterior fossa, compared with only 15% - 20% of adult tumors. For segmenting PF tumors we should have features to study the characteristics of tumors. In literature, different types of texture features such as Fractal Dimension (FD) and Multifractional Brownian Motion (mBm) have been exploited for measuring randomness associated with brain and tumor tissues structures, and the varying appearance of tissues in magnetic resonance images (MRI). For selecting best features techniques such as neural network and boosting methods have been exploited. However, neural network cannot descirbe about the properties of texture features. We explore methods such as information theroetic methods which can perform feature selection based on properties of texture features. The primary contribution of this dissertation is investigating efficacy of different image features such as intensity, fractal texture, and level - set shape in segmentation of PF tumor for pediatric patients. We explore effectiveness of using four different feature selection and three different segmentation techniques respectively to discriminate tumor regions from normal tissue in multimodal brain MRI. Our research suggest that Kullback - Leibler Divergence (KLD) measure for feature ranking and selection and Expectation Maximization (EM) algorithm for feature fusion and tumor segmentation offer the best performance for the patient data in this study. To improve segmentation accuracy, we need to consider abnormalities such as cyst, edema and necrosis which surround tumors. In this work, we exploit features which describe properties of cyst and technique which can be used to segment it. To achieve this goal, we extend the two class KLD techniques to multiclass feature selection techniques, so that we can effectively select features for tumor, cyst and non tumor tissues. We compute segemntation accuracy by computing number of pixels segemented to total number of pixels for the best features. For automated process we integrate the inhomoheneity correction, feature selection using KLD and segmentation in an integrated EM framework. To validate results we have used similarity coefficients for computing the robustness of segmented tumor and cyst.


Data is provided by the student.

Library Comment

Dissertation or thesis originally submitted to the local University of Memphis Electronic Theses & dissertation (ETD) Repository.