Table 3.

Performance achieved over the test set by each of the 48 considered model configurations

Arch.	Init.	Opt.	L.F.	Re	Pr	f1
MobileNet	R	Adadelta	CCE	0.797	0.802	0.796
			MSE	0.806	0.810	0.806
		Adam	CCE	0.831	0.841	0.832
			MSE	0.847	0.856	0.849
		RMSprop	CCE	0.781	0.802	0.782
			MSE	0.833	0.847	0.835
		SGD	CCE	0.764	0.778	0.767
			MSE	0.737	0.758	0.740
	I	Adadelta	CCE	0.877	0.881	0.878
			MSE	0.851	0.850	0.849
		Adam	CCE	0.851	0.856	0.852
			MSE	0.834	0.841	0.832
		RMSprop	CCE	0.838	0.849	0.839
			MSE	0.873	0.875	0.873
		SGD	CCE	0.797	0.805	0.798
			MSE	0.781	0.780	0.778
NASNet	R	Adadelta	CCE	0.802	0.803	0.801
			MSE	0.767	0.768	0.763
		Adam	CCE	0.804	0.817	0.804
			MSE	0.811	0.821	0.812
		RMSprop	CCE	0.752	0.751	0.749
			MSE	0.787	0.809	0.790
		SGD	CCE	0.794	0.792	0.792
			MSE	0.805	0.811	0.806
	I	Adadelta	CCE	0.836	0.837	0.835
			MSE	0.828	0.833	0.829
		Adam	CCE	0.838	0.840	0.837
			MSE	0.821	0.829	0.821
		RMSprop	CCE	0.795	0.795	0.793
			MSE	0.850	0.857	0.850
		SGD	CCE	0.811	0.813	0.811
			MSE	0.783	0.787	0.784
ResNet50	R	Adadelta	CCE	0.836	0.839	0.837
			MSE	0.828	0.829	0.827
		Adam	CCE	0.708	0.722	0.709
			MSE	0.699	0.704	0.697
		RMSprop	CCE	0.702	0.711	0.700
			MSE	0.711	0.740	0.715
		SGD	CCE	0.799	0.800	0.796
			MSE	0.807	0.805	0.805
	I	Adadelta	CCE	0.845	0.845	0.844
			MSE	0.855	0.856	0.855
		Adam	CCE	0.825	0.834	0.826
			MSE	0.813	0.815	0.813
		RMSprop	CCE	0.807	0.810	0.806
			MSE	0.827	0.831	0.826
		SGD	CCE	0.831	0.831	0.830
			MSE	0.781	0.791	0.779
VGG19	R	Adadelta	CCE	0.734	0.744	0.734
			MSE	0.822	0.826	0.822
		Adam	CCE	0.810	0.814	0.809
			MSE	0.751	0.768	0.751
		RMSprop	CCE	0.799	0.807	0.800
			MSE	0.791	0.803	0.793
		SGD	CCE	0.684	0.693	0.685
			MSE	0.819	0.821	0.818
	I	Adadelta	CCE	0.835	0.835	0.834
			MSE	0.806	0.808	0.805
		Adam	CCE	0.874	0.876	0.873
			MSE	0.885	0.887	0.884
		RMSprop	CCE	0.867	0.872	0.867
			MSE	0.833	0.842	0.832
		SGD	CCE	0.792	0.794	0.791
			MSE	0.753	0.769	0.754
InceptionV3	R	Adadelta	CCE	0.780	0.781	0.778
			MSE	0.764	0.773	0.763
		Adam	CCE	0.814	0.819	0.815
			MSE	0.757	0.766	0.755
		RMSprop	CCE	0.791	0.796	0.791
			MSE	0.793	0.796	0.792
		SGD	CCE	0.658	0.665	0.654
			MSE	0.719	0.731	0.717
	I	Adadelta	CCE	0.876	0.880	0.877
			MSE	0.837	0.844	0.838
		Adam	CCE	0.859	0.863	0.859
			MSE	0.835	0.838	0.836
		RMSprop	CCE	0.835	0.837	0.835
			MSE	0.848	0.850	0.848
		SGD	CCE	0.834	0.842	0.834
			MSE	0.836	0.838	0.836

The four leftmost columns account for the network meta-parameters, where specifically: Arch. stands for the deep network architecture, Init. stands for weight initialization (with R = random, I = from ImageNet), Opt. denotes the optimizer, and L.F. is the loss function (CCE = categorical croos-entropy, MSE = mean square error). The performance achieved by each model are in the three rightmost columns and are expressed in terms of Recall (Re), Precision (Pr) and f1. The model that achieves the highest value of f1 over the test set is highlighted in bold