Detection and Localization of Covid-19 on Chest Radiographs by Deep Learning Algorithms

The COVID-19 pandemic has caused one of the most severe global health crises in human history. Despite the progress made, the risk of contamination still exists due to new variants and insufficient booster vaccinations. Deep learning algorithms could be beneficial for the early diagnosis of Covid-19. This paper presents an application of deep learning techniques for detecting and localizing Covid-19 lung lesions on chest radiographs. It is a multi-class classification approach using three deep learning architectures: DensNet169, VGG16 and a non-pretrained sequential architecture. Indeed, transfer learning and ensemble learning techniques were used to classify the radiographs into three classes: “COVID”, “Pneumonia” and “Normal”. The database used consisted of 3225 chest radiographs selected by the team radiologist from the COVIDx - CXR version 8 database, one of the largest public databases. State of the art results were obtained for the three architectures and their different combinations, with “Accuracy” values above 83% for all architectures and reaching 96% for ensembles, which is consistent with results found in the literature. Class activation mapping (CAM) techniques were employed to localize and visualize the COVID lesions detected on chest radiographs.