Date of Award


Degree Name

Doctor of Philosophy


Electrical and Computer Engineering

First Advisor

Dr. Ikhlas M. Abdel-Qader

Second Advisor

Dr. Janos Grantner

Third Advisor

Dr. Azim Houshyar


Brain tumor, pattern recognition, feature extraction, MRI image, neural network, wavelet transform


A brain tumor is the most common disease that affects the central nervous system (CNS), the brain, and spinal cord. It can be diagnosed using the safest and most reliable imaging modality, the Magnetic Resonance Imaging (MRI), by radiologists who may use the assistance of computer-aided diagnosis (CAD) tools. Automated diagnosis is sought because it is essential to overcome the drawbacks of the manual diagnosis, such as time and the stress of viewing MRI images for long hours, and the human error potential. Image analysis and machine learning algorithms are tools that can be used to build an intelligent CAD system capable of analyzing brain tumors and formulating a diagnosis on its own. Hence, it is essential to design a CAD system that is capable of extracting meaningful and precise information, and rendering an error-free diagnosis. Consequently, many researchers have proposed different methods to develop a CAD system to detect and classify abnormal growths in brain images.

This dissertation presents a hybrid system for tumor classification from brain MRI images. The hybrid system is composed of a set of statistical-based features and deep neural networks. Segments of the MRI, from within the region of interest (ROI), are transformed into the two-dimensional Discrete Wavelet Transform and the two-dimensional Gabor filter methods. This allows the set of features to encompass all the directional information of the spatial domain tumor characteristics. A classifier system is developed using two types of neural network algorithms, Stacked Sparse Autoencoder (SSA) and Softmax Classifier. For the sparse autoencoder training, the sparsity regularization and L2-weight regularization are proposed. Sparsity regularization is used for its ability to control the firing of the neurons in the hidden layer, whereas L2-weight regularization is used for its ability to reduce the effect of overfitting. Two national brain tumor datasets were used to verify and validate the proposed system. The first dataset is a much larger dataset consisting of 3,064 slices of T1-weighted MRI with three kinds of tumors: Meningioma, Glioma, and Pituitary. The second dataset consists of 200 MRI slices with low-grade and high-grade Glioma tumors collected from the BRATS dataset. Implementation results using the first dataset achieved a total accuracy of 94.0%, and a specificity of 96.2%, 97.8%, and 97.3% for Meningioma, Glioma, and Pituitary tumors respectively. Using the second dataset, accuracy is at 98.8 %. Experimental results indicate not only that this system is effective, but also show that it outperforms the comparable methods.

Access Setting

Dissertation-Open Access