الفهرس 
Chapter 1 : IntroductionOnly 14 pages are availabe for public view
 |  | from | 144 |  |  |
| | | from | 144 | | |
Abstract
The brain tumor is a serious disease, and the number of people who are dying because of brain tumors is increasing. The utilization of different brain images has been expanding, which makes manually examining each image a labor-intensive task. The successful early diagnosis of brain tumors plays a major role in improving the treatment circumstances, diagnostic accuracy, reducing the required time and thus improving patient survival. Thus, there is a crucial need for computer-aided methods with better accuracy for early tumor diagnosis. Computer-aided brain tumor diagnosis from MRI images comprises tumor detection, segmentation, and classification.
In this thesis, new methods were investigated for accurate detection, segmentation, and classification of brain tumor using different databases. Method 1 presents a CAD system for classifying MRIs according to the presence or absence of tumors. The developed CAD system uses Kmeans, discrete wavelet transform (DWT), principal component analysis (PCA), and kernel support vector machine (KSVM) methods for segmentation, feature detection, feature reduction, and MRI classification, respectively. The experiments showed that the GRB kernel achieved a classification accuracy of 100%.
Method 2 is a CAD system for detecting the presence and the absence of a brain tumor and also classifying it into benign or malignant. This proposed CAD uses K-means, discrete wavelet transforms (DWT), principal component analysis (PCA), and kernel support vector machine (KSVM) methods for segmentation, feature detection, feature reduction, and MRI classification, respectively. The experiments showed that the proposed CAD system achieved a classification accuracy of 100%.
Method 3 is a two-phases multi-model deep learning-based system for brain tumor detection and localization from MRIs. The main goals of this study are to classify MRIs into normal and abnormal and accurately localize the tumor within the abnormal MRI images. The first system phase used CNN and ECOCSVM approach for feature extraction and classification, respectively. A five-layer R-CNN was used for tumor localization in the second system phase. The method achieved an accuracy of 99.55% and DICE score of 0.87.
Method 4 is a Two parallel U-Net with asymmetric residual-based deep convolutional neural networks (TPUAR-Nets) for brain tumor segmentation from MRI images. The TPUAR-Net model offers several advantages, including the possibility of considering both local and global features to learn both high-level and low-level features simultaneously. The deployment of the fully connected layer, the residual blocks, and the skip connection can overcome the vanishing gradient problem. The TPUAR-Net architecture achieved a maximum Dice score of 0.89.
Method 5 is the Deep convolutional neural network (DCNN) structure for brain tumor detection from MRI images. DCNN architecture has achieved an accuracy of 97.8%. Further improvements have been realized to improve the network by developing two parallel deep convolutional neural networks (PDCNN) structures for brain tumor, glioma in particular, detection and classification from MRI images in method 6, which takes the advantage of using both global and local features extracted from the two parallel stages. The PDCNN structure has achieved promising results of 97.44%, 98.0%, and 97.0% for accuracy, sensitivity, and specificity, respectively.