Table 3. Performance metrics achieved by the coarse models using the CheXpert test set.

The coarse models are initialized with ImageNet pretrained weights (conventional transfer learning) and trained end-to-end to learn CXR modality-specific weights using the CheXpert data set to classify the CXRs into normal and abnormal classes. The custom CNN is initialized with random weights. Data in parenthesis are 95% CI for the AUC values that were calculated to be the Wilson score interval which corresponds to separate 2-sided confidence intervals with individual coverage probabilities of sqrt(0.95).

Model	Accuracy	AUC	Sensitivity	Specificity	F measure	MCC
Custom CNN	0.8018	0.8898 (0.8813, 0.8983)	0.9030	0.5980	0.7952	0.5356
VGG-16	0.8904	0.9649(0.9599, 0.9699)	0.9173	0.8448	0.8904	0.7530
VGG-19	0.8799	0.9432 (0.9369, 0.9495)	0.9115	0.8165	0.8798	0.7288
Inception-V3	0.8835	0.9571 (0.9516, 0.9626)	0.9028	0.8363	0.8840	0.7402
Xception	0.8720	0.9401 (0.9337, 0.9465)	0.9005	0.8148	0.8723	0.7126
DenseNet-121	0.8839	0.9493 (0.9434, 0.9552)	0.9140	0.8233	0.8838	0.7378
MobileNet	0.8797	0.9456 (0.9395, 0.9517)	0.9073	0.8244	0.8799	0.7295
NASNet-mobile	0.8824	0.9552 (0.9496, 0.9608)	0.9045	0.8380	0.8828	0.7369

Notes.

Bold values indicate superior performance.