Table 4.

The comparison of models trained using different numbers of physician annotated images. The weakly supervised method uses only image-level annotations automatically produced from clinical diagnosis.

	Annotated images	AUROC	AUPRC
Supervised	N = 1040 (20%)	0.933	0.921
	N = 2081 (40%)	0.959	0.949
	N = 3122 (60%)	0.961	0.952
	N = 4163 (80%)	0.969	0.963
	N = 5204 (100%)	0.973	0.963
Weakly supervised	Image-level only	0.967	0.957

AUROC area under the receiver operating characteristic curve, AUPRC area under the precision-recall curve.