Fig. 5
From: Improving lesion detection in mammograms by leveraging a Cycle-GAN-based lesion remover
The ROC curves and associated AUCs of lesion detection networks on the test set. The test set included 507 recalled lesion and 987 normal tissue patches. Among deep network architectures considered in this study, ViT performed best over other architectures, regardless of its versions (Baseline, Highlighted, and Combined). We found the effectiveness of our proposed lesion highlighter for all architectures. Specifically, for ResNet18, both Highlighted and Combined versions performed better than its Baseline version (p < 0.0001, Table 1). For other more advanced and complex state-of-the-art networks, Combined versions performed better than their Baselines (p < 0.0001, Table 1)