COVision: Convolutional Neural Network for the Differentiation of COVID-19 from Common Pulmonary Conditions Using CT Scans

Abstract With the growing amount of COVID-19 cases, especially in developing countries with limited medical resources, it is essential to accurately diagnose COVID-19 with high specificity. Due to characteristic ground-glass opacities (GGOs), present in both COVID-19 and other acute lung diseases, misdiagnosis occurs often — 26.6% of the time in manual interpretations of CT scans. Current deep learning models can identify COVID-19 but cannot distinguish it from other common lung diseases like bacterial pneumonia. COVision is a multiclassification convolutional neural network (CNN) that can differentiate COVID-19 from other common lung diseases, with a low false-positivity rate. This CNN achieved an accuracy of 95.8%, AUROC of 0.970, and specificity of 98%. We found statistical significance that our CNN performs better than three independent radiologists with at least 10 years of experience, especially at differentiating COVID-19 from pneumonia. After training our CNN with 105,000 CT slices, we analyzed our CNN’s activation maps and found that lesions in COVID-19 presented peripherally, closer to the pleura, whereas pneumonia lesions presented centrally. Finally, using a federated averaging model, we ensemble our CNN with a pretrained clinical factors neural network (CFNN) to create a comprehensive diagnostic tool.