Comparison of the performances of deep learning models in classification of Covid-19 disease

The new coronavirus 2019 (Covid-2019), which first appeared in Wuhan, China in December 2019, spread rapidly in all countries and became a pandemic. It is very important to detect positive cases as early as possible in order to prevent the further spread of this epidemic, which has a devastating impact on both public health and the global economy in daily life, and to treat affected patients quickly. Radiology imaging techniques are used for artificial intelligence-based auxiliary diagnostic tools in the detection of Covid-19 disease. Lung X-ray images are one of the possible methods of detecting Covid-19. In this thesis, experimental studies were carried out within the framework of three scenarios with pre-trained deep learning models on two different datasets with Covid-19 diagnosis. The first dataset consists of two-classes computed tomography images of 1252 diagnosed with Covid-19 and 1229 undiagnosed. The second dataset includes Covid-19, pneumonia, and normal classes which have 2313 X-ray images in each class. In the first scenario, the original pre-trained deep learning models were used in the experiments, while in the second scenario, modified pre-trained models were used. Finally, experiments with traditional machine learning classifiers were conducted on the features extracted by original pre-trained models in the third scenario. In the experimental studies, higher accuracies were obtained with the changes made in the fully connected layers of the original pre-trained models compared to the other scenarios, and among these models, D_EfficientNetV2B0 provided 98,99% and 97,45% accuracies in two-class and three-class datasets, respectively.