Table 2.

Performance of deep learning model and two experienced radiologists in training and internal validation cohort

	Results (n)				Test performance (%)
	TP	TN	FP	FN	AUC [95%CI]	Sensitivity [95%CI]	Specificity [95%CI]	Accuracy [95%CI]	P †
Training	1368	5051	113	696	82.05 [81.01-83.08]	66.28 (1368/2064) [64.10-68.32]	97.81 (5051/5164) [97.41-98.20]	88.81 (6419/7228) [88.06-89.53]
Internal Validation
Deep-learning model	356	1268	23	160	83.61 [81.58-85.64]	68.99 (356/516) [65.12-73.06]	98.22 (1268/1291) [97.44-98.92]	89.87 (1624/1807) [88.39-91.23]
Radiologist 1	181	1239	52	335	65.52 [63.40-67.65]	35.08 (181/516) [30.81-39.34]	95.97 (1239/1291) [94.89-96.98]	78.58 (1420/1807) [76.62-80.45]	< 0.0001
Radiologist 2	115	1248	43	401	59.48 [57.62-61.34]	22.29 (115/516) [18.60-25.78]	96.67 (1248/1291) [95.66-97.60]	75.43 (1363/1807) [73.38-77.40]	< 0.0001

TP = true positive, TN = true negative, FP = false positive, FN = false negative

†: compare between radiologists and deep learning model.

Delong's test was used to compare the AUCs.