Fig. 2.

Examples of valid (left) and invalid (right) CXR scans. A CNN-based classifier was trained and used to automatically filter outliers; only valid PA-view CXRs of adults were retained for labeling.