Table 1:

Comparison of performance of GMIC and the baselines on NYUBCS. For both GMIC and ResNet-34, we reported test AUC (mean and standard deviation) of top-5 models that achieved highest validation AUC in identifying breasts with malignant findings. We also measure the total number of learnable parameters in millions, peak GPU memory usage (Mem) for training a single exam (4 images), time taken for forward (Fwd) and backward (Bwd) propagation in milliseconds, and number of floating-point operations (FLOPs) in billions.

Model	AUC(M)	AUC(B)	#Param	Mem(GB)	Fwd/Bwd (ms)	FLOPs
ResNet-34	0.736 ± 0.026	0.684 ± 0.015	21.30M	13.95	189/459	1622B
ResNet-34-1 × 1 conv	0.889 ± 0.015	0.772 ± 0.008	21.30M	12.58	201/450	1625B
DMV-CNN (w/o heatmaps)	0.827 ± 0.008	0.731 ± 0.004	6.13M	2.4	38/86	65B
DMV-CNN (w/ heatmaps)	0.886 ± 0.003	0.747 ± 0.002	6.13M	2.4	38/86	65B
Faster R-CNN	0.908 ± 0.014	0.761 ± 0.008	104.8M	25.75	920/2019	-3
GMIC-ResNet-18	0.913 ± 0.007	0.791 ± 0.005	15.17M	3.01	46/82	122B
GMIC-ResNet-34	0.909 ± 0.005	0.790 ± 0.006	25.29M	3.45	58/94	180B
GMIC-ResNet-50	0.915 ± 0.005	0.797 ± 0.003	27.95M	5.05	66/131	194B
GMIC-ResNet-18-ensemble	0.930	0.800	-	-	-	-
GMIC-ResNet-34-ensemble	0.920	0.795	-	-	-	-
GMIC-ResNet-50-ensemble	0.927	0.805	-	-	-	-