مشخصات پژوهش

In this study, the DiagCovidPNA method is proposed to aid in diagnosing and differentiating COVID-19 disease and viral and bacterial pneumonia (PNA) from chest X-ray (CXR) images. The proposed method is based on application-specific convolutional neural network (CNN) architectures. DiagCovidPNA is a four-class (i.e., COVID-19, normal lung, bacterial PNA, and viral PNA) diagnostic method that is offered in two different systems: 1) DiagCovidPNA and 2) Hybrid DiagCovidPNA. In the first system, a single CNN is used to classify the input images into four classes. In the second system, to further differentiate bacterial and viral PNAs more, the input images are classified into four classes using a combination of a three-class (i.e., COVID-19, normal lung, PNA) CNN-based classifier and a two-class (i.e., bacterial PNA, and viral PNA) CNN-based classifier. A comprehensive labeled database of CXR images for training and testing purposes has been compiled using two sources of available images. In addition, the transfer learning method is applied to reduce the issue of insufficient training data. Extensive experiments are conducted to evaluate and compare both systems to existing methods. According to the results, the diagnostic accuracy of DiagCovidPNA and Hybrid DiagCovidPNA is 95% and 97.50%, respectively. Consequently, Hybrid DiagCovidPNA improves accuracy by approximately 8% compared to other deep learning-based methods with a same structure and approximately 3% compared to the parallel transfer deep learning-based method. All DiagCovidPNA implementation source codes and the compiled CXR database are publicly available to the research community.