COVID-19 detection from chest x-ray images and hybrid model recommendation with convolutional neural networks

The COVID-19 pandemic, which emerged at the end of 2019, continues to be effective. In the first days of the epidemic, no medicine was found against the virus. Indicators such as the variability of the effectiveness of passive and mRNA-based vaccines against different mutations of the virus, the unknown possible side effects, the presence of unvaccinated people in the population, possible problems with the supply and storage of the vaccine show that the epidemic will not end in the near future. Despite the World Health Organization's recommendations to combat the epidemic, countries follow different methods against the virus. However, despite all the differences, the common method followed is the detection and isolation of the disease. That's why the correct diagnosis of coronavirus is critical. Considering the fact that there are many countries in the world that still cannot reach test kits and the unique difficulties of establishing a new infrastructure, it is very important that the solution to be put forward uses existing medical capabilities. X-Ray imaging devices, which is accepted as standard equipment in health centers, seems promising for this issue since it is cost effective and easily accessible. Within the scope of the thesis, the detection of COVID-19 from lung X-Ray images was studied. Corona virus detection, which can be confused with other lung diseases from X-Ray images, can only be made by expert radiologists. It seems possible to overcome the related problems with convolutional neural networks that can classify with high accuracy. Ensembled systems, that is, architectures with multiple classifiers, show much more effective results in real-world problems of health and computer vision. In these architectures, while a reduction in total system error is achieved, more successful classifications are made. Within the scope of the related study, binary and multiple classification performances were examined using MobileNetV2, DenseNet121, InceptionResNetV2 and Xception networks. In the next step, these models are combined with ensemble learning and a hybrid model is developed for more effective binary and multiple classification.