Table 1.

Performance results of vision transformer-based transfer learning models (ViT-TL), vision transformer models trained from scratch (ViT-scratch), and convolutional neural network-based transfer learning models (CNN-TL) approaches.

Approach	Model	Accuracy	F1 score	Precision	Recall	AUC
ViT-TL	ViTB-16	0.854 ± 0.01	0.844 ± 0.01	0.872 ± 0.01	0.826 ± 0.01	0.870 ± 0.01
	ViTB-32	0.842 ± 0.02	0.830 ± 0.03	0.840 ± 0.03	0.822 ± 0.03	0.865 ± 0.02
	ViTL-32	0.810 ± 0.02	0.796 ± 0.02	0.814 ± 0.02	0.784 ± 0.02	0.845 ± 0.02
ViT-scratch	ViTB-16	0.624 ± 0.02	0.604 ± 0.03	0.618 ± 0.03	0.600 ± 0.04	0.666 ± 0.01
	ViTB-32	0.622 ± 0.02	0.596 ± 0.03	0.594 ± 0.03	0.584 ± 0.04	0.648 ± 0.01
	ViTL-32	0.600 ± 0.02	0.590 ± 0.03	0.584 ± 0.03	0.572 ± 0.04	0.644 ± 0.01
CNN-TL	ResNet50	0.772 ± 0.02	0.756 ± 0.02	0.804 ± 0.02	0.726 ± 0.03	0.785 ± 0.02
	EfficientNetB2	0.680 ± 0.07	0.608 ± 0.12	0.674 ± 0.07	0.614 ± 0.10	0.744 ± 0.06
	InceptionV3	0.804 ± 0.02	0.766 ± 0.02	0.848 ± 0.02	0.722 ± 0.03	0.823 ± 0.01