Deep Learning-based Detection for COVID-19 from Chest CT using Weak Label

Chuansheng Zheng; Xianbo Deng; Qing Fu; Qiang Zhou; Jiapei Feng; Hui Ma; Wenyu Liu; Xinggang Wang

. CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity. For every testing CT scan, we used the trained DeCoVNet to predict its probability of COVID-19. By comparing with their binary ground-truth labels, we plotted ROC and PR curves as shown in Fig. 3 and Fig. 4 respectively. In the ROC, we obtained a ROC AUC value of 0.959. When true positive rate (TPR, i.e., sensitivity) was approximately 0.95, our model obtained a true negative rate (TNR, i.e., specificity) of 0.786; when TNR was approximately 0.95, our model obtained a TPR of 0.880; there was another operating showed that our algorithm obtained both TPR and FPR larger than 0.9, i.e., sensitivity=0.907 and specificity=0.911. On the PR curve, our model obtained a PR AUC of 0.975.