Figure 4.

Results of the deep learning architecture trained on CheXpert for seven radiological findings (a–g) and one radiological diagnosis (h) on a separate 234-cases test dataset, selected at random within the 500 test set studies of the CheXpert dataset (cf. Irvin et al. for details). The latter was composed of randomly sampled studies from the full dataset with no patient overlap. Three board-certified radiologists individually annotated each of the studies in this test set.