Fig. 3

One-vs-others receiver operating characteristic (ROC) plots of our deep neural network (DNN) tested in the internal (a) and external (b) test sets are presented. Overall, the model showed acceptable generalizability for Imitator and tuberculosis (TB) predictions between the two tests. While the model was best at predicting non-tuberculous mycobacteria (NTM) in the internal cohort, it achieved the worst result in external cohort. This finding might come from great heterogeneity between NTM groups in the internal and external test sets. c, d demonstrates confusion matrices of DNN’s performance (c) and the pooled performance of the 12 pulmonologists (d) on the internal test set. The major distinction between human experts and machines can be found in NTM prediction. Even though the recruited pulmonologists are experts of mycobacterial diseases, they tended to make random guesses when chest X-rays (CXRs) of NTM were presented to them