Table 3.

Performance comparison of models based on different initial weights and learning rates for different ratios of the training dataset.

Ratio	Learning rate	AUC
		Scratch-based model	ImageNet-based model	CAMELYON-based model
20%	5e−2	0.516	0.511	0.572
	5e−3	0.623	0.766	0.799
	5e−4	0.887	0.894	0.912
	5e−5	0.890	0.891	0.920
	5e−6	0.813	0.872	0.910
40%	5e−2	0.546	0.599	0.597
	5e−3	0.759	0.803	0.812
	5e−4	0.897	0.919	0.929
	5e−5	0.881	0.905	0.930
	5e−6	0.822	0.824	0.931
100%	5e−2	0.546	0.616	0.631
	5e−3	0.754	0.804	0.811
	5e−4	0.914	0.943	0.944
	5e−5	0.921	0.944	0.938
	5e−6	0.911	0.906	0.901

AUC was measured for patch-level evaluation in AMC dataset.