الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 109 |  |  |
| | | from | 109 | | |
Abstract
Medical diagnosis from CT scans and X-ray images have a critical role in the early diagnosis of Pneumonia. Medical diagnosis has a considerable potential in improving the survival rate of patients. COVID-19 causes the deadliest Pneumonia in the world. COVID-19 diagnosis is still a time-consuming and long awaiting process. On the other hand, the number of cases of Pneumonia that affect people is increasing day by day due to mutation changes according to environmental pollution, neglectance of quarantine, lack of medical care, and so forth. Therefore, CT and X-ray chest screening has started to be used globally. However, analyzing medical images is a severe load for radiologists. In the third-world countries, medical diagnosis is dramatically expensive. Therefore, automatic diagnosis is relatively easy and dramatically practical to help in COVID-19 detection with higher accuracy.Convolutional Neural Networks (CNN) represent a promising solution in the early diagnosis from medical images. However, free access to classified medical images presents a dilemma. Therefore, the best solution is to transfer learning (TL), where knowledge is shared from realistic images to medical images to eliminate the need for so many images. However, TL is not a promising solution in diagnosing Pneumonia and COVID-19 because of the enormous similarity in symptoms and the considerable overlap in severity between medical images. Therefore, we resort to using Single-Image Super-Resolution (SISR) as a pre-stage before the classification process. The SISR is one of the best ways to increase image quality and enhance perceptual image consistency.In this thesis, we concentrate on the quality enhancement of medical images for effective automated medical diagnosis, especially in detecting COVID-19. Quality is improved by enlarging the image dimensions with a technique known as Super Resolution (SR). Generative Adversarial Network (GAN) algorithms are used to enlarge images. The super-resolution stage is used as a pre-processing stage to improve the performance of the automated diagnosis process. The well-known GAN algorithm is used to generate images from noise. It was used in various applications, including generating High-Resolution (HR) images from Low-Resolution (LR) images. The GAN consists of two sub-models, the first is generative and the second is discriminatory. Each model tries to deceive the other, the generator generates an HR image, and the discriminator decides if it is HR or LR.Content loss is used as a loss function in GAN. Content loss depends on the concept of Mean Squared Error (MSE). However, instead of comparing the generated image to the reference image. We use feature extraction from two images by DenseNet 121 model. Then, we apply the MSE criterion on the extracted features. The DenseNet 121 model is first trained to classify Pneumonia images, and then the feature extraction part is cropped and used in a loss function.
The subsequent image classification process is implemented with DL models. The models comprise a designed one from scratch and 14 pre-train models with the help of TL. The objective is to use pre-training model weights with some fine-tuning to avoid the need for a large training database. The proposed framework comprises both SR and classification, and it is superior in performance to that working on LR images. The proposed network achieves a classification accuracy of 99%, which is superior to the recent traditional algorithms.