Deep ConvNets for COVID-19 Recognition from Chest X-Rays

Abstract

Coronavirus disease 2019 (COVID-19) is an extremely contagious and quickly spreading Coronavirus infestation. Severe Acute Respiratory Syndrome (SARS)-CoV and Middle East Respiratory Syndrome (MERS)-CoV, which outbreak in 2002 and 2011, and the current COVID-19 pandemic are all from the same family of coronavirus. The fatality rate due to SARS and MERS was higher than COVID-19. However, the spread of those was limited to few countries, while COVID19 affected more than 200 countries worldwide, causing over 3 million causalities and infected more than 145million people as of April 25, 2021. Given the effects of COVID-19 on pulmonary tissues, chest radiographic imaging has become a necessity for screening and monitoring the disease. Recently, numerous studies have proposed Deep Learning approaches based on Convolutional Neural Networks (CNNs, or ConvNets) for the automatic diagnosis of COVID-19 from chest X-rays (CXR). Although these methods achieved astonishing performance in early detection and diagnosis, they have used limited CXR repositories for evaluation, usually with a few hundred COVID-19 CXR images only. Thus, such data scarcity prevents reliable evaluation with the potential of overfitting. In addition, manual annotation of X-rays (delineation of the lung, or infection regions) is another challenge due to the extensive time and manual labor required from the physicians. Therefore, most of the proposed studies showed no or limited performance in infection localization and severi grading of COVID-19 pneumonia.In this thesis, in order to overcome the aforementioned limitations and challenges, we have conducted the following: (i) compiled the largest COVID-19 benchmark dataset, namely COVID-QU, which consists of 11,956 COVID-19, 11,263 non-COVID-19, 10,701 normal, 134 SARS, and 144 MERS CXR images, (ii) generated ground-truth lung segmentation masks for the entire COVID-QU dataset using an elegant human-machine collaborative approach, (iii) proposed a systematic approach to segment the lung, detect, localize, and quantify COVID-19 infections from CXR images, (iv) Trained and evaluated the proposed system for lung segmentation, infection segmentation, and two classification tasks: ?) COVID-19 detection from the predecessor COVID family members, SARS, and MERS, ?) COVID-19 detection from non-COVID-19 infections, and normal cases. A detailed set of experiments using several state-of-the-art ConvNets showed top performance for the lung segmentation task with Intersection over Union (IoU) of 96.11% and Dice Similarity Coefficient (DSC) of 97.99%. Besides, COVID-19 infections of various shapes and types were reliably localized with 83.05% IoU and 88.21% DSC. Moreover, the proposed system was able to discriminate between different COVID family members, which is an extremely challenging task for medical doctors without the aid of clinical data. Sensitivities of 96.94%, 79.68%, and 90.26% were achieved for classifying COVID-19, MERS, and SARS classes, respectively. Furthermore, a good performance was obtained for the second classification scheme with sensitivities of 91.52%, 93.21%, and 91.12 for COVID 19, non-COVID, and normal classes, respectively.