Table 2:

Classification performance on CBIS-DDSM. For GMIC, we reported test AUC of top-5 models that achieved highest validation AUC in identifying breasts with malignant findings. We compared GMIC with five baselines. The performance of Deep MIL, RGP, and GGP in this table was originally reported in Shu et al. (2020).

Model	AUC(M)
ResNet-34	0.792 ± 0.014
ResNet-34-1×1 conv	0.800 ± 0.011
Deep MIL (Zhu et al., 2017)	0.791 ± 0.0002
RGP (Shu et al., 2020)	0.838 ± 0.0001
GGP (Shu et al., 2020)	0.823 ± 0.0002
GMIC-ResNet-18	0.833 ± 0.004
GMIC-ResNet-18 (best)	0.840
GMIC-ResNet-34	0.830 ± 0.003
GMIC-ResNet-50	0.828 ± 0.001
GMIC-ResNet-18-ensemble	0.858
GMIC-ResNet-34-ensemble	0.849
GMIC-ResNet-50-ensemble	0.849