الفهرس | Only 14 pages are availabe for public view |
Abstract Biomedical data analysis is playing an important role in health-care provision services. Recently, there has been a signicant rise in the amount of biomedical data collection. As the amount of biomedical data produced every day increases, the risk of human making analytical and diagnostic mistakes also increases. Hence, there is a need for the development of fast, accurate, reliable and automatic means for analysis of biomedical data. To overcome these limitations, this thesis concerns two case studies. In the rst case study, an Electroencephalogram (EEG) signals for epileptic seizure diagnosis are considered. In the second case, two studies are conducted using X-ray, Computed Tomography (CT) and Ultrasonic (US) images for lung disease diagnosis. In the rst case study, two proposed approaches are introduced for epileptic seizure diagnosis. The rst proposed approach depends on Machine Learning (ML) and some statistical features in Empirical Mode Decomposition (EMD) domain. The simulation results reveal the eectiveness of the proposed approach from the accuracy perspective, but the approach becomes complex as the number of features increase. In order to overcome this limitation an endto- end classication approach is introduced. The proposed approach depends on spectrogram and Phase Space Reconstruction (PSR) as a pre-processing stage to convert the 1-D EEG signals into 2-D images. Dierent Convolutional Neural Network (CNN) models are used with this strategy. The simulation results prove that the PSR gives better classication results than those given by the spectrogram. The PSR is a direct projection from the time domain, which keeps the main trend of dierent signal activities. In the second case study, ML and dierent deep learning strategies are introduced for COVID-19 detection from X-ray, CT and US images. First of all, ML is adopted on features extracted manually from images. Twelve classiers are compared for this task. To extend the feasibility of this study, we have modied the features extraction strategy to give deep features. Four pre-trained models are adopted in this study. Moreover, Transfer Learning (TL) is also introduced in this study to enhance the COVID-19 detection process. The selected classication hierarchy is also compared with a CNN model built from scratch to prove its quality of classication. Simulation results prove that deep features and TL methods provide the highest accuracy of 100%. In addition, an automated approach for diagnosis of lung diseases from chest X-ray and CT images is introduced. This approach depends on Superresolution (SR) as a tool for enhancing the details of chest X-ray and CT images. Deep Learning (DL) algorithms are exploited for the SR task. In addition, the classication process is performed also with DL. Moreover, a comparison study between softmax and Multi-class Support Vector Machine (MCSVM) classiers is presented. The proposed approach is evaluated using three dierent publicly available X-ray and CT images. Simulation results reveal that the combination between image SR and InceptionResNetv2 gives the best classication accuracy of 98.028%. |