Table 2.

Comparison of the verification performance on two different subsets of the ChestX-ray14 dataset that either contain foreign material or not (first two rows). Furthermore, we show the verification results for the CheXpert dataset and the COVID-19 Image Data Collection (last two rows). We present the AUC (together with the lower and upper bounds of the 95% confidence intervals from 10,000 bootstrap runs), the accuracy, the specificity, the recall, the precision, and the F1-score.

Dataset	Subset	AUC + 95 % CI	Accuracy ($\frac{TP+TN}{P+N}$)	Specificity ($\frac{\mathit{TN}}{N}$)	Recall ($\frac{\mathit{TP}}{P}$)	Precision ($\frac{\mathit{TP}}{TP+FP}$)	F1-score
ChestX-ray14	w/ foreign material	$0.{9970}_{0.9938}^{0.9993}$	$0.9796(\frac{672}{686})$	$0.9854(\frac{338}{343})$	$0.9738(\frac{334}{343})$	$0.9853(\frac{334}{339})$	0.9795
	w/o foreign material	$0.{9972}_{0.9909}^{0.9999}$	$0.9862(\frac{430}{436})$	$0.9908(\frac{216}{218})$	$0.9817(\frac{214}{218})$	$0.9907(\frac{214}{216})$	0.9862
CheXpert	–	$0.{9870}_{0.9855}^{0.9884}$	$0.9440(\frac{15,562}{16,486})$	$0.9629(\frac{7,937}{8,243})$	$0.9250(\frac{7,625}{8,243})$	$0.9614(\frac{7,625}{7,931})$	0.9429
COVID-19	–	$0.{9763}_{0.9696}^{0.9825}$	$0.9180(\frac{1,421}{1,548})$	$0.9780(\frac{757}{774})$	$0.8579(\frac{664}{774})$	$0.9750(\frac{664}{681})$	0.9127