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

Abstract

Abstract
Background: Quick and precise identification of people suspected of having COVID-19 plays a key function in imposing quarantine at the right time and providing medical treatment, and results not only in societal benefits but also helps in the development of an improved health system. Building a deep-learning framework for automated identification 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 and localization of swellings.

Methods: In all cases, subjects’ chest areas were segmented by means of a pre-trained U-Net; the segmented 3D chest areas were submitted as inputs to a 3D deep neural network to forecast the likelihood of infection with COVID-19; the swellings were restricted by joining the initiation areas within the classification system and the unsupervised linked elements. A total of 499 3D CT scans were utilized for training worldwide and 131 3D CT scans were utilized for verification.

Results: The algorithm took only 1.93 seconds to process the CT amount of a single affected person using a special graphics processing unit (GPU). Interesting results were obtained in terms of the development of societal challenges and better health policy.

Conclusions: The deep-learning model can precisely forecast COVID-19 infectious probabilities and detect swelling areas in chest CT, with no requirement for training swellings. The easy-to-train and high-functioning deep-learning algorithm offers a fast method to classify people affected by COVID-19, which is useful to monitor the 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