Table 2.
Dataset | Phase 1(P1) | Phase 2(P2) | Phase 3(P3) | MIMIC-CXR filtered (M) |
---|---|---|---|---|
# cases | 295 | 250 | 2,507 | 194,495 |
# cases studies with eye-tracking data | 285 | 240 | 2,507 | 0 |
# MIMIC-CXR images | 59 | 50 | 2,507 | |
# subjects | 58 | 50 | 2,110 | 60,018 |
% female | 63.8 | 54.0 | 50.7 | 53.9 |
% male | 36.2 | 46.0 | 49.1 | 45.7 |
% test set | 15.3 | 14.0 | 20.2 | 1.4 |
% Normal Radiograph (P1, P2, P3) & No Finding (M) | 18.0 | 24.4 | 22.8 | 32.9 |
% Abnormal mediastinal contour (P2,P3) & Wide mediastinum (P1) | 2.7 | 5.6 | 2.7 | |
% Acute fracture (P2,P3) & Fracture (P1, M) | 5.1 | 2.8 | 1.0 | 1.9 |
% Airway wall thickening (P1) | 7.1 | |||
% Atelectasis (P1,P2,P3,M) | 41.4 | 27.6 | 25.8 | 20.5 |
% Cardiomegaly (M) | 19.8 | |||
% Consolidation (P1,P2,P3,M) | 28.5 | 28.8 | 25.9 | 4.7 |
% Enlarged cardiac silhouette (P1,P2,P3) | 28.1 | 28.4 | 21.8 | |
% Enlarged Cardiomediastinum (M) | 3.2 | |||
% Enlarged hilum (P2,P3) | 2.8 | 1.9 | ||
% Groundglass opacity (P1,P2,P3) | 9.2 | 18.8 | 12.6 | |
% Hiatal hernia (P2,P3) | 0.0 | 0.9 | ||
% High lung volume/emphysema (P2,P3) & Emphysema (P1) | 3.1 | 3.2 | 2.9 | |
% Interstitial lung disease (P2,P3) & Fibrosis (P1) | 1.7 | 1.2 | 1.0 | |
% Lung nodule or mass (P2,P3) & Lung Lesion (M) | 1.6 | 5.1 | 2.7 | |
% Lung Opacity (M) | 22.8 | |||
% Mass (P1) | 0.7 | |||
% Nodule (P1) | 4.7 | |||
% Other (P1,P2,P3) | 13.9 | 8.8 | 6.0 | |
% Pleural abnormality (P2,P3) | 30.0 | 29.5 | ||
% Pleural Effusion (P1,M) | 31.2 | 24.2 | ||
% Pleural thickening (P1) | 2.0 | |||
% Pleural Other (M) | 0.9 | |||
% Pneumonia (M) | 7.2 | |||
% Pneumothorax (P1,P2,P3,M) | 4.7 | 4.4 | 2.9 | 4.6 |
% Pulmonary edema (P1,P2,P3) & Edema (M) | 13.9 | 13.6 | 13.7 | 12.1 |
% Quality issue (P1) | 3.4 | |||
% Support devices (P1,P2,P3,M) | 36.9 | 34.8 | 44.8 | 29.3 |
The dataset where each label was present is shown inside parentheses. “Normal radiograph” represents CXRs for which no other label was selected. Table cells are left blank for labels that were not present in that dataset. For how the labels of the different datasets are related, check Fig. 5.