Fig 1.
Flow chart of the study population.
Table 1.
Demographic description of the dataset.
Table 2.
Lesion types of referable thoracic abnormalities on chest radiographs (determined based on computed tomography [CT]).
Fig 2.
The prevalence of normal cases, and target and non-target lesions of a deep-learning algorithm (DLA) showing significant differences between the three institutions.
Institution G has fewer normal cases and more DLA-non-target lesions compared to those of the other two institutions.
Table 3.
Standalone performance of the deep-learning algorithm (DLA) for visible referable thoracic abnormalities on chest radiographs in the multicenter health screening cohort based on chest computed tomography (CT) findings.
Fig 3.
Receiver operating characteristic curve (ROC) curve of a deep-learning algorithm (DLA) for referable thoracic abnormalities on chest radiography based on different standard reference methods.
The area under the ROC curve (AUC) shows better performance when using visible and clear visible CXR compared to using CT as ground truth methods, except for institution G (C).