Table 1: Data used to train and evaluate the pneumonia segmentation model.

Convolutional neural networks were trained on publically-available frontal chest radiographs and radiologist-defined bounding boxes demarcating areas of lung parenchyma associated with pneumonia. 22,000 radiographs were used for model training and the remaining were reserved to evaluate performance.

	Total	Training	Validation
N (%)	25,684	22,000 (85.6%)	3,684 (14.4%)
% Male	56.8%	56.6%	57.9%
Mean Age, years (range)	47 (1–92)	47 (1–92)	46.9 (3–91)
% AP	45.6%	45.4%	46.7%
N (%) Pneumonia	5,656 (22.0)	4,796 (21.8)	860 (23.3)
N (%) Abnormal, not Pneumonia	11,512 (44.8)	9,878 (44.8)	1,634 (44.4)
N (%) Normal	8,516 (33.2)	7,326 (33.3)	1,190 (32.3)