Table 1.

The averaged performance results over 5-fold cross validation of the proposed multistage transfer learning and its comparison against conventional transfer learning. TL: transfer learning; CNN: convolutional neural network; AUC: area under ROC curve, Avg.: average; CTL: conventional transfer learning; MSTL: multistage transfer learning; SGD: stochastic gradient descent.

TL Type	CNN	Optimizer	AUC	F1 Measure	Specificity	Sensitivity	Loss	Test Accuracy (%)	Avg. Test Acc. (%)
CTL method	InceptionV3	SGD	0.903	0.833	0.87	0.80	0.412	83.50 ± 5.491	83
		Adam	0.778	0.605	0.66	0.75	9.570	70.50 ± 6.085
		Adagrad	0.976	0.967	1	0.93	0.195	96.49 ± 2.091
	EfficientNetb2	SGD	0.717	0.664	0.71	0.61	0.644	66.00 ± 1.895	83
		Adam	0.993	0.948	0.98	0.90	0.194	93.99 ± 4.726
		Adagrad	0.980	0.904	0.98	0.81	0.300	89.50 ± 2.709
	ResNet50	SGD	0.960	0.902	0.90	0.91	0.296	90.50 ± 2.850	89
		Adam	0.817	0.698	0.66	0.97	0.117	81.50 ± 10.216
		Adagrad	0.989	0.974	0.97	0.98	0.084	97.50 ± 2.165
The proposed MSTL method	InceptionV3	SGD	0.935	0.873	0.83	0.94	0.458	88.50 ± 3.758	92
		Adam	0.967	0.930	0.94	0.92	0.292	93.00 ± 2.291
		Adagrad	0.981	0.945	0.95	0.94	0.208	94.50 ± 0.935
	EfficientNetB2	SGD	0.820	0.762	0.77	0.76	0.606	76.50 ± 3.409	90
		Adam	0.998	0.980	0.98	0.98	0.067	97.99 ± 1.249
		Adagrad	0.992	0.965	0.97	0.96	0.207	96.50 ± 1.274
	ResNet50	SGD	0.995	0.985	0.99	0.98	0.065	98.50 ± 1.118	98
		Adam	0.986	0.964	0.96	0.97	0.216	96.49 ± 1.000
		Adagrad	0.999	0.989	0.98	1	0.030	99.00 ± 0.612