Diagnosis of COVID-19 Disease Using Convolutional Neural Network Models Based Transfer Learning

COVID-19 disease is similar to normal pneumonia caused by bacteria or other viruses. Therefore, the manual classification of lung diseases is very hard to discover, particularly the distinction between COVID-19 and NON-COVID-19 disease. COVID-19 causes infections on one or both lungs which appear as inflammations across lung cells. This can lead to dangerous complications that might cause death in the case of gaining or having an immune disease. The problem of COVID-19 is that its symptoms are similar to conventional chest respiratory diseases like flu disease and chest pain while breathing or coughing produces mucus, high fever, absence of appetite, abdominal pain, vomiting, and diarrhea. In most cases, a deep manual analysis of the chest’s X-ray or computed tomography (CT) image can lead to an authentic diagnosis of COVID-19. Otherwise, manual analysis is not sufficient to distinguish between pneumonia and COVID-19 disease. Thus, specialists need additional expensive tools to confirm their initial hypothesis or diagnosis using real-time polymerase chain reaction (RT-PCR) test or MRI imaging. However, a traditional diagnosis of COVID-19 or other pneumonia takes a lot of time from specialists, which is so significant parameter in the case of a pandemic, whereas, a lot of patients are surcharging hospital services. In such a case, an automatic method for analyzing x-ray chest images is needed. In this regard, the research work has taken advantage of proposing a convolutional neural network method for COVID-19 and pneumonia classification. The X-ray processing have been chosen as a diagnosis way because of its availability in hospitals as a cheap imaging tool compared to other technologies. In this work, three CNN models based on VGG-16, VGG19, and MobileNet were trained using the zero-shot transfer learning technique. The best results are obtained on VGG-19 based model: 96.97% accuracy, 100% precision, 100% F1-score, and 99% recall.