Development and Evaluation of an AI System for COVID-19

Cheng Jin; Weixiang Chen; Yukun Cao; Zhanwei Xu; Xin Zhang; Lei Deng; Chuansheng Zheng; Jie Zhou; Heshui Shi; Jianjiang Feng

in different subsets (Figure 3 b, c). We found that the different performances of different subsets were highly correlated to the number of abnormal slices, which means smaller lesion with fewer abnormal slices are more difficult to diagnose ( Part of the patients in the database have multi-stage CTs. We compared the diagnostic performance of stage I and stage II and fusion of stage I, II in the external test cohort (Figure 3 a, b, d) . The experiment suggested that the performance of the AI system is independent of the progress of the disease because of no significant differences between performances of different stages. The statistical results also showed that fusion of stage I and II could slightly improve the performance of diagnosis, in which the fusion method we adopted is to simply average the scores of two stages. We did not test more complex fusion methods which may overestimate the performance since each negative case has 8 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.