الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 119 |  |  |
| | | from | 119 | | |
Abstract
This thesis is mainly concerned with Electroencephalography (EEG) signal analysis. Both seizure detection and perdition are considered from the records of epilepsy patients in this thesis. The meaning of detection is to determine the seizure onset as early as possible from its beginning using machine or deep learning techniques rather than manual inspection by specialists. On the other hand, seizure prediction is performed to give alerts on expected seizure states prior to occurrence as early as possible. This action is very important for patients and care givers, while the seizure detection is important for long-term analysis of EEG signal activities. Three proposals are presented in this thesis for EEG seizure detection and prediction. The first proposal depends on the transformation of signal segments into 2D matrices, which are processed with Scale-Invariant Feature Transform (SIFT) algorithm. The SIFT is well-known in image processing for its ability to determine feature points. This property is exploited to localize important points and signal segments and their numbers. The number of important points is statistically used to discriminate between seizure and normal activities in seizure detection scenarios. The second proposal depends on the application of wavelet transform on EEG signal attributes in the wavelet domain for EEG seizure prediction. The third proposal depends on extracting spectrograms of EEG signals using them as images for discrimination between different signal activities. A deep convolutional neural network is used for the classification process. Both two-state and three-state classification scenarios are considered for the seizure detection and prediction processes. The presented scenarios are compared with traditional algorithms. Their results reveal superiority compared to those of traditional algorithms.