Differentiating COVID-19 from other types of pneumonia with convolutional neural networks

Abstract INTRODUCTION A widely-used method for diagnosing COVID-19 is the nucleic acid test based on real-time reverse transcriptase-polymerase chain reaction (RT-PCR). However, the sensitivity of real time RT-PCR tests is low and it can take up to 8 hours to receive the test results. Radiologic methods can provide higher sensitivity. The aim of this study is to investigate the use of X-ray and convolutional neural networks for the diagnosis of COVID-19 and to differentiate it from viral and/or bacterial pneumonia, as 2-class (bacterial pneumonia vs COVID-19 and viral pneumonia vs COVID-19) and 3- class (bacterial pneumonia, COVID-19, and healthy group (BCH), and among viral pneumonia, COVID- 19, and healthy group (VCH)) experiments. METHODS 225 COVID-19, 1,583 healthy control, 2,780 bacterial pneumonia, and 1,493 viral pneumonia chest X-ray images were used. 2-class- and 3-class-experiments were performed with different convolutional neural network (ConvNet) architectures, with different variations of convolutional layers and fully-connected layers. RESULTS The results showed that bacterial pneumonia vs COVID-19 and viral pneumonia vs COVID- 19 reached a mean ROC AUC of 97.32% and 96.80%, respectively. In the 3-class-experiments, macro-average F1 scores of 95.79% and 94.59% were obtained in terms of detecting COVID-19 among BCH and VCH, respectively. CONCLUSIONS The ConvNet was able to distinguish the COVID-19 images among non-COVID-19 images, namely bacterial and viral pneumonia as well as normal X-ray images.