. Author manuscript; available in PMC: 2021 Jan 1.

Published in final edited form as: Med Phys. 2019 Nov 19;47(1):110–118. doi: 10.1002/mp.13886

Table 2.

Comparison of the deep learning model performance for predicting short-term breast cancer risk. The numbers are average AUCs of 10-fold cross-validation with 95% Confidence Interval [CI]. The highest AUC was 0.73 and 0.72 when using the whole-breast and dense tissue region, respectively, both achieved by the GoogLeNet-LDA model.

Model	CC view		MLO view		MLO + CC view
Model	Whole–breast	Dense tissue	Whole–breast	Dense tissue	Whole–breast	Dense tissue
End-to-End GoogLeNet	0.68(95% CI:0.60 – 0.75)	0.64(95% CI:0.55–0.72)	0.60(95% CI:0.55–0.64)	0.62(95% CI:0.53–0.72)	0.62(95% CI:0.58–0.66)	0.67(95% CI:0.61–0.73)
GoogLeNet-LDA	0.73(95% CI:0.68 – 0.78)	0.70(95% CI:0.65–0.76)	0.69(95% CI:0.65–0.72)	0.67(95% CI:0.59–0.75)	0.64(95% CI:0.58– 0.70)	0.72(95% CI:0.67– 0.76)