TABLE II. Performances of Different Models.

Evaluation Metric		Clinical model	Radiomic model	Pure DenseNet	DenseNet + Seg	De-COVID19-Net
Training Set	AUC	0.740 [0.649, 0.829]	0.865 [0.807, 0.922]	0.874 [0.824, 0.923]	0.922 [0.887, 0.957]	0.952 [0.928, 0.977]
	Accuracy J0.5	0.752	0.821	0.780	0.793	0.882
	Sensitivity J0.5	0.630	0.783	0.848	0.935	0.957
	Specificity J0.5	0.780	0.830	0.765	0.760	0.865
	Accuracy J0.6	0.626	0,772	0.756	0.744	0.850
	Sensitivity J0.6	0.804	0.848	0.848	1.000	1.000
	Specificity J0.6	0.585	0.755	0.735	0.685	0.815
Test Set	AUC	0.733 [0.608, 0.855]	0.850 [0.771, 0.927]	0.870 [0.802, 0.937]	0.906 [0.852, 0.959]	0.943 [0.904, 0.981]
	Accuracy J0.5	0.717	0.783	0.742	0.783	0.875
	Sensitivity J0.5	0.542	0.708	0.792	0.917	0.917
	Specificity J0.5	0.760	0.802	0.729	0.750	0.864
	Accuracy J0.6	0.633	0.775	0.717	0.725	0.842
	Sensitivity J0.6	0.708	0.792	0.833	1.000	1.00
	Specificity J0.6	0.615	0.s771	0.688	0.656	0.802

The thresholds to calculate accuracy, sensitivity, and specificity were obtained by the weighted Youden's Index . The AUC confidence intervals were obtained by a 2000-time bootstrap. Our proposed De-COVID19-Net achieved the best performance in all metrics (in bold). The construction of the models is illustrated in Section II. E.