Özet:

In today's conditions, although the pandemic has lost its effects and is in the process of ending, COVID-19 still shows its effects on people as mild. With the developments in image processing and Artificial Intelligence technologies, the correct detection of such viruses in the early stages will both help the healing process of the disease quickly by applying the right treatment, and will alleviate the burden on health systems. In this study, it has been tried to create a high-accuracy and reliable model that tries to distinguish COVID-19 and viral pneumonia diseases from chest X-ray images. For this purpose, a comprehensive modeling study has been carried out by applying the AlexNet and GoogleNet special architectures of convolutional neural networks, which are deep learning algorithms, directly to their original versions and to their improved versions with transfer learning. The data set used in the modeling process (COVID-19 Radiography Database) is a popular data set and an unbalanced data set with 3 classes and a different number of samples in each class. By applying data reduction and increase methods to this data set, 2 new balanced data sets containing equal number of samples in each class were created. By dividing the original dataset and newly created datasets into training and test datasets at a rate of 80-20, and also by cross validating 3, 5, and 10 times, model performances were measured and the model with the best performance was tried to be found. As a result, the best model was found with 99.90% accuracy, as the data set balanced by data augmentation method was divided according to cross validation 10 times and AlexNet architecture developed with transfer learning was applied.