Table 5.

Performance metrics of the image-level malignancy prediction model with different lesion identification models (one-class, malignancy and sub-types) and aggregation strategies (average, noisy OR and max pooling). The best performing combination is highlighted in boldface.

	Average		Noisy OR		Max
Detection	AUC	AP	AUC	AP	AUC	AP
Malignancy	0.7698	0.7759	0.7698	0.7461	0.7690	0.7772
Sub-type	0.7784	0.7822	0.7698	0.7600	0.7643	0.7705
One-class	0.7759	0.7650	0.7783	0.7661	0.7868	0.7821