Fig 1.
Flow-chart diagram illustrating development, validation, and testing of DL algorithm.
Fig 2.
ROC curves for the DL algorithm and the four test radiologists (R1, R2, R3 and R4) in for pulmonary opacities (O), pleural effusion (PE), hilar prominence (HP) and enlarged cardiac silhouette (C).
Fig 3.
True positive pulmonary opacities.
Frontal CXR belonging to two separate patients. Unprocessed CXR (a, c) demonstrate nodular opacities in the right lung. The corresponding heat maps from DL algorithm (b, d) accurately detected and annotated (in red) these abnormalities.
Table 1.
Summary of AUC for detection of radiographic abnormalities in CXR.
Fig 4.
Frontal chest radiograph (a) without radiographic abnormality. The DL algorithm generated heat map (b) labels false positive pleural effusion on the right side.
Fig 5.
Implanted Port catheter projects in the left mid lung zone (a). Heat map from DL algorithm misinterpreted the port catheter as a focal pulmonary opacity. This is apparent on the accompanying heat map image (b).
Table 2.
Summary of AUC for change in abnormalities over follow-up CXR.
Fig 6.
Frontal CXR demonstrates a subtle patchy opacity in right mid zone (a), which is annotated (in red) on the heat map generated from DL algorithm (b). On subsequent follow-up CXR (c), the heat map did not mark any abnormality (resolution).