Table 7:

Ablation study: effect of different choice of aggregation function. We report the performance achieved by parameterizing f_agg as global average pooling (GAP), global maximum pooling (GMP), and top t% pooling. For each setting, we trained five GMIC-ResNet-18 models and report the mean and standard deviation of AUC and DSC.

fagg	AUC(M)	AUC(B)	DSC(M)	DSC(B)
GMP	0.890 ± 0.02	0.785 ± 0.012	0.127 ± 0.052	0.103 ± 0.060
t = 1%	0.906 ± 0.01	0.784 ± 0.007	0.190 ± 0.030	0.147 ± 0.053
t = 2%	0.916 ± 0.009	0.790 ± 0.007	0.203 ± 0.013	0.191 ± 0.042
t = 3%	0.913 ± 0.007	0.791 ± 0.004	0.228 ± 0.036	0.178 ± 0.041
t = 5%	0.912 ± 0.009	0.790 ± 0.002	0.172 ± 0.004	0.194 ± 0.027
t = 10%	0.914 ± 0.005	0.791 ± 0.008	0.156 ± 0.050	0.182 ± 0.028
t = 20%	0.907 ± 0.017	0.785 ± 0.008	0.126 ± 0.048	0.182 ± 0.040
GAP	0.903 ± 0.02	0.783 ± 0.012	0.065 ± 0.006	0.181 ± 0.011