Table 2.

Readers’ and model’s evaluation at the mammogram level

	True positives	True negatives	Weighted accuracya	Cohen κ
Model versus Reader 1	26/32 81% (67–94%)	97/127 76% (68–84%)	0.027 + 0.813 0.840 (0.693–0.966)	0.45 (0.30–0.59)
Model versus Reader 2	18/23 78% (60–94%)	98/136 72% (64–80%)	0.025 + 0.783 0.808 (0.623–0.972)	0.31 (0.17–0.46)
Reader 2 versus Reader 1	17/32 53% (35–70%)	121/127 95% (91–98%)	0.033 + 0.530 0.563 (0.387–0.737)	0.54 (0.35–0.70)
Model versus consensus	26/32 78% (67–83%)	97/127 75% (67–83%)	0.25 + 0.579 0.829 (0.680–1.000)	0.32 (0.17–0.49)

^a Class weights are assigned inversely proportional to their respective frequencies as assigned by Reader 1. The 95% CI values are indicated in parentheses. All p-values are below 0.001