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

Published in final edited form as: Med Image Anal. 2020 Dec 16;68:101908. doi: 10.1016/j.media.2020.101908

Table 5:

Ablation study: effectiveness of incorporating both global and local features. We report the mean and standard deviation of the test AUC for top-5 GMIC-ResNet-18. We experimented with 4 GMIC variants that use ŷ_global, ŷ_local, the average of ŷ_global and ŷ_local, and ŷ_fusion as predictions. The proposed design that uses ŷ_fusion as predictions outperforms all variants.

Prediction	AUC(M)	AUC(B)
${\hat{y}}_{global}$	0.892 ± 0.009	0.776 ± 0.004
${\hat{y}}_{local}$	0.897 ± 0.004	0.778 ± 0.005
$\frac{1}{2} ({\hat{y}}_{local} + {\hat{y}}_{global})$	0.905 ± 0.006	0.785 ± 0.004
${\hat{y}}_{fusion}$	0.913 ± 0.007	0.791 ± 0.005