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 motivation of this study was to utilize AI to alleviate the problem of shortage of professional interpretations for CT images when the epidemic is still fast spreading. Though there were many effective applications of medical AI in previous studies [13, 23] , developing AI for automatic COVID-19 detection was still a challenging task. Firstly, in the current emergency situation, the number of enrolled patients is relatively smaller compared with that used in previous studies [13, 23] ; and patients enrolled in our study were clinically diagnosed cases with COVID-19, because the majority of them did not undergo the nucleic acid testing due to the sudden outbreak and limited medical resource in such a short time period. Secondly, the lesions of COVID-19 in CT volumes were not labeled by radiologists and only patientlevel labels (i.e., COVID-positive or COVID-negative) were utilized for training the AI algorithm in our study. Thirdly, some small infected areas of COVID-19 have the potential to be missed even by professional radiologists, and whether it is feasible to be detected by deep learning-based 3D DCNN model remains unclear. We hypothesized to solve these problems by proposing a delicate 3D DCNN, i.e., DeCoVNet. It solved the first problem by applying extensive data augmentation on training CT volumes to obtain more training examples. The second problem was solved by regarding the COVID-19 detection problem as a weakly-supervised learning problem [24] , i.e., detecting COVID-19 without annotating the regions of COVID-19 lesions. In the designed DeCoVNet, we used the spatially global pooling layer and the temporally global pooling layer to technically handle the weakly-supervised COVID-19 detection problem. The third problem was addressed by taking the advantages of deep learning and utilizing a pre-trained UNet for providing the lung masks to guide the learning of DeCoVNet.