Automatic Diagnosis of Pneumonia and COVID-19 Using Convolutional Neural Networks and Transfer Learning Amina Bekkouche, Mohammed Merzoug, Fethallah Hadjila, Ismail Bellaouedj, Abdelhak Etchiali Computer Science Departement, Abou bekr Belkaid University of Tlemcen B.P 119 Faculty of Sciences, Tlemcen, 13000 Algeria Abstract- Several studies are currently ex- ploring the diagnosis of lung disorders using deep learning analysis of medical images. Deep learn- ing is also considered to be a valuable aid to experts in the interpretation of medical images. Heuristics such as transfer learning are becom- ing more common; these methods (based on pre- trained models) are utilized as the basis for com- puter vision tasks and can significantly improve various issues. This work proposes models built on Convolutional Neural Networks (CNNs) that incorporate transfer learning to identify various pneumonia infections in X-ray images. The ex- periments show that the model based on Xcep- tion network outperforms many existing state-of- the-art methods and several recent backbones. Keywords- Convolutional neural networks, COVID-19 detection, Deep learning, Transfer learning, Pneumonia diagnosis, XceptionNet, In- ception modules, Residual networks. I. Introduction Pneumonia diseases are among the most common and severe infectious diseases that exhibit high morbidity and mortality, particularly in the elderly and immuno- compromised. The diagnosis of several pulmonary illnesses, including COVID-19, lung opacity, viral pneu- monia, and bacterial pneumonia, could greatly benefit from recent Al-based technologies, when integrated into the medical field. In general, the problem of diagnosing depends on the symptoms, mostly through physical examinations and other tests such as chest X-ray images. The implementation of an automatic diagnosing sys- tem based on the latest algorithms will constitute a great assistance for doctors. The challenge for AI (Artificial In- telligence) in the healthcare field is how to execute tasks that require the highest level of cognitive functions, and how to ensure the reduction of bias during the learning process. In addition, what matters the most is data, which represent the fuel of any AI algorithm. The uti- lized data should be of high quality and homogeneity degree and must be relevant to the actual subject. In this paper, we focus on the automatic diagnosis of pul- monary diseases (and especially COVID-19) using chest X-ray images from ”COVID-19 Radiography Dataset” while improving the accuracy. In this context, we pro- pose and compare six types of CNNs (Convectional Neu- ral Network) that use transfer learning and customized dense layers to detect the presence of COVID-19 and other pneumonia in X-ray images. Our proposed models can be summarized as follows: • A backbone based on MobileNet V2 plus a single dense layers. • A backbone based on Inception V2 plus three dense layers. • A backbone based on ResNet50 V2 plus three dense layers. • A backbone based on Xception plus a single dense layers. • A backbone based on Inception ResNet V2 plus three dense layers. • A backbone based on DenseNet plus a single dense layers. The remainder of the paper is structured as follows. The state-of-art with different sub-classes of COVID-19 diagnosis approaches is presented in Section II. The pro- posed method with the different used backbones is de- scribed in Section III. The results are shown and dis- cussed in Section IV. In the final section V., we present conclusions and future directions. International Journal of Neural Networks and Advanced Applications DOI: 10.46300/91016.2022.9.7 Volume 9, 2022 E-ISSN: 2313-0563 40