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

The workflow of deep-learning based diagnosis model is shown in Figure 1 b. CT cases were firstly divided to different cohorts and extracted to slices since our model takes 2D slices as input. Then after slice level training, our model can accurately predict whether the input slices come from COVID-19 subjects. With a top-k average block, our model finally fused slice results into case level diagnosis. The model was implemented in 2D not only because 2D network was easily to train with more training samples, but also because slice-level scores can be used for abnormal slice locating. We fine-tuned our diagnosis model on a training dataset consisting of normal and abnormal slices from COVID-19 positive cases and obtained the abnormal slice locating model. Other parts of our system are described in Methods.