Comparing the performance of ResNets on COVID-19 diagnosis using CT scans

Abstract

To the date, the Coronavirus disease 2019 (COVID-19) is still spreading around the world. Reverse transcription polymerase chain reaction (RT-PCR), which is regarded as the gold standard by the anti-epidemic front. Unfortunately, due to its lack of sensitivity (only 60%~71%), unstability, long processing time, some alternative diagnostic tools including computed tomography (CT) have been developed using convolutional neural networks (CNNs). This paper retrospectively collected 3227 CT scans (including 1601 COVID-19 positive and 1626 negative samples), on which ResNet-18 outperforms all three classical ResNets with the precision of 0.971, accuracy of 0.943, F1 of 0.942, sensitivity of 0.914, specificity of 0.973 and AUC of 0.985 respectively. Finally, the visualization technique classification activation map (CAM) was applied for lesions localization.