Table 2.

Diagnostic performance of binary classifiers for diseases versus normal conditions built using the original images and cropped images, where the numbers of test images are as listed in Table 1.

Type	Binary classifier	Performance
	Category	Accuracy	Sensitivity	Specificity
Original images	Bronchiolitis	85.84%	89.21%	80.46%
		(0.8029–0.8982)	(0.8273–0.9394)	(0.6980–0.8750)
	Bronchopneumonia	86.38%	88.82%	81.61%
		(0.8054–0.8988)	(0.8323–0.9321)	(0.7210–.8915)
	Lobar pneumonia	93.99%	94.79%	93.10%
		(0.8962–0.9627)	(0.8823–0.9796)	(0.8509–0.9749)
	Pneumothorax	92.25%	80.95%	97.26%
		(0.8529–0.9535)	(0.6600–0.9111)	(0.9212–1.0000)
Cropped images	Bronchiolitis	87.50%	89.21%	84.71%
		(0.8259–0.9063)	(0.8280–0.9328)	(0.7590–0.9162)
	Bronchopneumonia	90.55%	91.72%	88.24%
		(0.8622–0.9331)	(0.8710–0.9545)	(0.8049–0.9412)
	Lobar pneumonia	96.69%	96.88%	96.47%
		(0.9194–0.9834)	(0.9145–0.9900)	(0.9065–0.9884)
	Pneumothorax	94.49%	90.48%	96.47%
		(0.8818–0.9685)	(0.9026–0.9892)	(0.8922–0.9886)