Table 2

Further tuning of the selected model (NASNetLarge) and the performance metrics

		New model specifications
	Original model (as in table 1)	horizontal_flip=True//epochs=25 (stopped at 12 epochs)	horizontal_flip=True//zoom_range=0.30//epochs=25 (stopped at 15 epochs)	decay=0.1/10//epochs=25 (stopped at 12 epochs)	decay=0.1/10//factor=0.3//epochs=25 (stopped at 12 epochs)
Loss, validation	0.526799	0.534797	0.537553	0.532246	0.532246
Accuracy, validation	0.742046	0.737500	0.739773	0.732386	0.732386
Loss, test	0.539906	0.550286	0.528204	0.538252	0.538252
Accuracy, test	0.731818	0.719318	0.735795	0.725568	0.725568
Precision, label 0 (moderate)	0.82	0.85	0.76	0.83	0.83
Recall, label 0 (moderate)	0.76	0.71	0.87	0.74	0.74
F1 score, label 0 (moderate)	0.79	0.77	0.81	0.78	0.78
Precision, label 1 (extreme)	0.60	0.57	0.66	0.59	0.59
Recall, label 1 (extreme)	0.68	0.75	0.47	0.71	0.71
F1 score, label 1 (extreme)	0.64	0.65	0.55	0.64	0.64

Green colour highlights the best metric, yellow colour highlights the second best metric and red colour highlights the third best metric row-wise considering only the new model specifications. The precision, recall and F1 score are presented as proportions (multiply by 100 to have percentages). receiver operating characteristic curves for each model are available in online supplemental materials.