Development and Evaluation of an AI System for COVID-19 Diagnosis

TLDR

Early detection of COVID‑19 using chest CT can enable timely treatment and help control spread, but the rapid increase in CT volumes outpaces the availability of human experts. The study proposes an AI system for rapid COVID‑19 diagnosis with accuracy comparable to experienced radiologists. The system was trained on 970 CT volumes from 496 confirmed COVID‑19 patients and 260 negative cases from three Wuhan hospitals, plus 1,125 negative cases from two public datasets, using a deep convolutional neural network. On an independent external verification set of 1,255 cases, the AI achieved 94.98% accuracy, 97.91% AUC, 94.06% sensitivity, and 95.47% specificity, performed comparably to five radiologists in a reader study, and was two orders of magnitude faster, with code available at the provided GitHub link.

Abstract

Abstract Early detection of COVID-19 based on chest CT will enable timely treatment of patients and help control the spread of the disease. With rapid spreading of COVID-19 in many countries, however, CT volumes of suspicious patients are increasing at a speed much faster than the availability of human experts. Here, we propose an artificial intelligence (AI) system for fast COVID-19 diagnosis with an accuracy comparable to experienced radiologists. A large dataset was constructed by collecting 970 CT volumes of 496 patients with confirmed COVID-19 and 260 negative cases from three hospitals in Wuhan, China, and 1,125 negative cases from two publicly available chest CT datasets. Trained using only 312 cases, our diagnosis system, which is based on deep convolutional neural network, is able to achieve an accuracy of 94.98%, an area under the receiver operating characteristic curve (AUC) of 97.91%, a sensitivity of 94.06%, and a specificity of 95.47% on an independent external verification dataset of 1,255 cases. In a reader study involving five radiologists, only one radiologist is slightly more accurate than the AI system. The AI system is two orders of magnitude faster than radiologists and the code is available at https://github.com/ChenWWWeixiang/diagnosis_covid19 .