Supervised framework for COVID-19 classification and lesion localization from chest CT

Abstract

AbstractBackground: Quick and precise identification of people suspected of having COVID-19 plays a key function inimposing quarantine at the right time and providing medical treatment, and results not only in societal benefits butalso helps in the development of an improved health system. Building a deep-learning framework for automatedidentification of COVID-19 using chest computed tomography (CT) is beneficial in tackling the epidemic.Aim: To outline a novel deep-learning model created using 3D CT volumes for COVID-19 classification andlocalization of swellings. Methods: In all cases, subjects’ chest areas were segmented by means of a pre-trained U-Net; the segmented 3Dchest areas were submitted as inputs to a 3D deep neural network to forecast the likelihood of infection withCOVID-19; the swellings were restricted by joining the initiation areas within the classification system and theunsupervised linked elements. A total of 499 3D CT scans were utilized for training worldwide and 131 3D CTscans were utilized for verification. Results: The algorithm took only 1.93 seconds to process the CT amount of a single affected person using aspecial graphics processing unit (GPU). Interesting results were obtained in terms of the development of societalchallenges and better health policy. Conclusions: The deep-learning model can precisely forecast COVID-19 infectious probabilities and detectswelling areas in chest CT, with no requirement for training swellings. The easy-to-train and high-functioningdeep-learning algorithm offers a fast method to classify people affected by COVID-19, which is useful to monitorthe SARS-CoV-2 epidemic. [Ethiop. J. Health Dev. 2020; 34(4):235-242] Key words: COVID-19, CT scan, deep learning, neural network, DeCoVNet, RT-PCR, computed tomography