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

The deep learning-based COVID-19 diagnostic algorithm used in our study is effective compared to recent deep learning-based computer-aided diagnosis methods. On the task of predicting the risk of lung cancer [13] , the deep learning model was trained on 42290 CT cases from 14851 patients and obtained 0.944 ROC AUC. On the task of critical findings from head CT [23] , the deep learning model was trained on 310055 head CT scans and obtained ROC AUC of 0.920. In our study, only 499 scans were used for training, but the obtained ROC AUC was 0.959. By comparing the data between them, it was able to find that the task of COVID-19 detection may be easier and the proposed deep learning algorithm was very powerful. As for the erroneous 12 false negative predictions in our results, the most possible explanations after we rechecked the original CT images were listed as follows: those lesions were slightly increased in CT densities, and images of those ground-glass opacities were very faint without consolidation.