Skip to main content
. 2020 Aug 14;117(35):21373–21380. doi: 10.1073/pnas.2003026117

Fig. 3.

Fig. 3.

The performance of deep CNN and other methods. We used Deep CNN, CytoDx, CellCNN, and FlowSOM to diagnose latent CMV in the test dataset. (A) The performances of the models measured by the receiver-operator characteristics curves. (B) The areas under the receiver-operator characteristics curves. The error bars represent the SD. The SD and P values are measured by 1,000-fold bootstrapping. (C) The performances of the models measured by the precision-recall curves. (D) The areas under the precision-recall curves. The error bars represent the SD. The SD and P values are measured by 1,000-fold bootstrapping. *P < 0.05; **P < 0.01.