Table 7.

This table summarises the results acquired from the fivefold stratified CV.

	Balanced accuracy	Macro precision	Macro recall	Macro F1 score	Kappa score	Inference time (s)	Training time (h)
EfficientNetB0	0.9303 ± 0.0322	0.9161 ± 0.0408	0.9303 ± 0.0322	0.9211 ± 0.0378	0.9180 ± 0.0362	0.0810 ± 0.0006	0.5565 ± 0.0088
EfficientNetV2B0	0.9076 ± 0.0398	0.8988 ± 0.0429	0.9076 ± 0.0398	0.9000 ± 0.0416	0.9040 ± 0.0455	0.0753 ± 0.0004	0.5630 ± 0.0287
EfficientNetV2B0-21k	0.9666 ± 0.0185	0.9646 ± 0.0174	0.9666 ± 0.0185	0.9642 ± 0.0184	0.9678 ± 0.0154	0.0758 ± 0.0001	0.5592 ± 0.0162
ResNetV1-50	0.9253 ± 0.0310	0.9244 ± 0.0358	0.9253 ± 0.0310	0.9206 ± 0.0334	0.9255 ± 0.0313	0.2184 ± 0.0014	0.5795 ± 0.0556
ResNetV2-50	0.9346 ± 0.0156	0.9199 ± 0.0276	0.9346 ± 0.0156	0.9259 ± 0.0202	0.9233 ± 0.0247	0.2277 ± 0.0010	0.5968 ± 0.0478
MobileNetV1	0.9518 ± 0.0232	0.9526 ± 0.0180	0.9518 ± 0.0232	0.9506 ± 0.0214	0.9543 ± 0.0181	0.0424 ± 0.0004	0.5628 ± 0.0340
MobileNetV2	0.9362 ± 0.0322	0.9339 ± 0.0251	0.9362 ± 0.0322	0.9314 ± 0.0305	0.9357 ± 0.0278	0.0456 ± 0.0011	0.5659 ± 0.0818

Each performance metric was reported in average (± standard deviation) form. The bold values represent the best score in each category. The best overall performing model was found to be EfficientNetV2-B0-21k, and the fastest CNN model was MobileNetV1.