Table 4.

Performance of the different deep learning models on the 80-pixel sub-database testing set. The best-achieved results are bold. For the ensemble learning models, WA stands for weighted averaging; UA stands for unweighted averaging, and MV stands for majority voting, and the 3 and 5 at the end of the ensemble models refer to top 3 or 5 base models. All metrics are measured in % unit.

Model	Accuracy	AUC	Precision	Recall	Specificity	F1-Score
MobileNet	95.82	95.73	94.90	95.15	96.30	95.02
MobileNetV2	95.29	94.87	96.36	92.26	97.48	94.27
EfficientNetB0	96.47	96.46	95.26	96.39	96.53	95.82
EfficientNetB1	96.50	96.41	95.83	95.83	96.99	95.83
DenseNet121	96.61	96.30	97.42	94.41	98.20	95.89
DenseNet169	96.67	96.70	95.26	96.88	96.52	96.07
InceptionV3	94.56	94.55	92.71	94.47	94.63	93.58
Xception	95.48	95.40	94.34	94.92	95.88	94.63
Ensemble-WA3	97.56	97.51	96.97	97.22	97.80	97.09
Ensemble-WA5	97.69	97.59	97.54	96.95	98.23	97.24
Ensemble-UA3	97.59	97.57	96.80	97.47	97.67	97.13
Ensemble-UA5	97.72	97.65	97.39	97.18	98.12	97.28
Ensemble-MV3	97.49	97.47	96.66	97.38	97.57	97.02
Ensemble-MV5	97.66	97.59	97.32	97.10	98.07	97.21