الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 133 |  |  |
| | | from | 133 | | |
Abstract
Worldwide, brain tumors are considered the main cause of mortality. Hence, an urgent need arises for a fully automated brain tumor detection system. Machine learning advancement inspired researchers to develop computer-aided diagnosis systems-based deep learning to decrease the dependency on the commonly used error-prone manual diagnosis. Generally, brain tumors can be categorized into meningioma, glioma, and pituitary. Accurate categorization of such tumors is vital for treatment planning and prognosis. To tackle this challenge, this thesis employed the deep learning (DL) network’s ability to generalize amidst diverse data. The proposed models offer an accurate multi-class brain tumor classification based on the convolutional neural network (CNN) architecture. The proposed multi-class brain tumor classification process was empowered by DL techniques, primarily the CNNs. In this thesis, three models based on CNN were introduced to enhance classification accuracy, reduce the computational cost, reduce the processing time, and reduce the overfitting problem. The proposed model 1 introduces an effective classification method designed to differentiate between different perspectives of three brain tumor types using images captured using magnetic resonance imaging (MRI). Model 1 comprises 10 layers, encompassing convolution layers, 1×1 convolution layers, average pooling, fully connected layers, and a SoftMax layer. Five iterations with transfer learning (TL), and five-fold cross-validation for retraining were employed to enhance the model’s performance. The proposed model 1 achieved an average accuracy of 98.63% over five iterations using TL and 98.86% through retrained five-fold crossvalidation (intra-fold TL). Then, proposed model 2 relies on combining iv the extracted features from three distinct CNN branches based on model 1. Each branch receives input from various transform domains of the original MRI, including the Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), and the time-domain image. The used CNN in model 2 included a concatenation layer, flatten layer, and dense layer before implementing the SoftMax layer. The results demonstrated the superiority of model 2 by its remarkable performance with 98.89% accuracy, 98.78% F1-score, 98.74% precision, 98.82% recall, and 99.44% specificity using 5-fold cross-validation. Finally, to solve the overfitting problem, model 3 was proposed using two paths of the CNN by inputting the original images to one path, and the DWT of the original image to the other path. Dropout layers were used to alleviate the overfitting risk, thereby improving the model’s generalizability. Additionally, multiple classifiers were employed to increase the model’s performance. The results of model 3 demonstrated accuracy of 99.51%. The accompanying high sensitivity and specificity metrics affirm the model’s ability to distinguish the three tumor categories. The model underwent validation using a separate test dataset, to confirm its ability to be generalized and used practically.