Table 4.

Normalized confusion matrix for the “real-world” dataset with the consensus decision of the two readers as ground truth, applying the four classes of the ACR density definition

	Reference:consensus decision	PREDICTED
		RMLO projections				RCC projections
		ACR A	ACR B	ACR C	ACR D	ACR A	ACR B	ACR C	ACR D
GROUND TRUTH	ACR A	73.7(89.5%)[89.5%]	26.3(10.5%)[10.5%]	0.0%(0.0%)[0.0%]	0.0%(0.0%)[0.0%]	79.0%(100.0%) [94.7% ]	21.1%(0.0%)[5.3%]	0.0%(0.0%)[0.0%]	0.0%(0.0%)[0.0%]
	ACR B	0.0%(18.2%)[9.1%]	93.9%(66.7%)[87.9%]	6.1%(15.2%)[3.0%]	0.0%(0.0%)[0.0%]	0.0% (6.9%) [20.7%]	93.1%(93.1%)[79.3%]	6.9%(0.0%)[0.0%]	0.0%(0.0%)[0.0%]
	ACR C	0.0%(0.0%)[0.0%]	16.0%(8.0%)[0.0%]	80.0%(84.0%)[100.0%]	4.0%(8.0%)[0.0%]	0.0% (0.0%) [0.0%]	8.1%(5.4%)[2.7%]	83.8%(94.6%)[75.7%]	8.1%(0.0%)[21.6%]
	ACR D	0.0%(0.0%)[0.0%]	0.0%(0.0%)[0.0%]	8.7%(26.1%)[8.7%]	91.3%(73.9%)[91.3%]	0.0% (0.0%) [0.0%]	0.0%(0.0%)[0.0%]	6.7%(46.7%)[0.0%]	93.3%(53.3%)[100.0%]

ACR, American College of Radiology ; dCNN, deep convolutional neural network;RCC, right craniocaudal;

Reader 1, no brackets; Reader 2, round brackets ( ); the dCNN, square brackets [ ].