Improved deep learning model for differentiating novel coronavirus pneumonia and influenza pneumonia

Min Zhou; Yong Chen; Dexiang Wang; Yanping Xu; Weiwu Yao; Jingwen Huang; Xiaoyan Jin; Zilai Pan; Jingwen Tan; Lan Wang; Yihan Xia; Longkuan Zou; Xin Xu; Jingqi Wei; Mingxin Guan; Jianxing Feng; Huan Zhang; Jieming Qu

The copyright holder for this preprint . https://doi.org/10.1101/2020.03.24.20043117 doi: medRxiv preprint measure, the Trinary scheme achieves score 0.847, which is higher than the Plain scheme (0.774) and also much higher than the specialist group (average 0.785) and the resident group (average 0.644). Trinary scheme was better correlated with specialists in both categories (Table E9 ). More details were in the Result E7. Table 1 summarized the CT devices on which both schemes and ten radiologists made wrong classification on cases from the external validation set. We first observed that the IP cases are from 10 CT devices, despite the fact that they were from the same center. The majority of the tested IP cases have been correctly classified by both schemes. The only exception was uCT 528, a new CT device. On uCT 528, eight patients were examined, seven of them were IP from Center 1, from which six and five cases were misclassified by Plain and Trinary scheme respectively. Yet more than three patients were also misdiagnosed by the specialists. The main manifestations were peripheral single or multiple ground grass opacities with or without patchy consolidation in the lower lobe or bilateral distribution, which mimic the findings of NCP, leading to the misclassification ( Figure E8 ). Another one was from Center 8 which was an NCP but misdiagnosed by all specialists ( Figure 4A ). The Trinary scheme performs better than the Plain scheme in this situation. Table 1 , the Plain scheme misclassified 16 cases and the Trinary scheme reduced it to 10. The error rates of both schemes for the two CT devices (SOMATOM Definition Flash and LightSpeed VCT) from Center 6 were exceptionally high compared to all other CT devices. As both of them only contributed IP training cases for the model, the classification model may learn the device specific features, and wrongly treated these features as specific to IP during training. During testing, the schemes would therefore tend to wrongly classify NCP.