Skip to main content

Table 1 Detection performances of various CNN architectures on non-highlighted and highlighted versions, and their differences

From: Improving lesion detection in mammograms by leveraging a Cycle-GAN-based lesion remover

Model

Type

AUC on Test set

Diff. over base [95% CI]

p-value

ResNet18

Baseline

0.914

  

N/A

Highlighted

 

0.963

0.049 [0.036, 0.062]

 < 0.0001*

Combined

 

0.963

0.050 [0.038, 0.061]

 < 0.0001*

DenseNet201

Baseline

0.966

  

N/A

Highlighted

 

0.969

0.003 [-0.004, 0.010]

0.426

Combined

 

0.974

0.008 [0.004, 0.012]

 < 0.001*

EfficientNetV2

Baseline

0.96

  

N/A

Highlighted

 

0.963

0.003 [-0.005, 0.011]

0.476

Combined

 

0.968

0.008 [0.004, 0.012]

 < 0.001*

ViT

Baseline

0.967

  

N/A

Highlighted

 

0.969

0.002 [-0.005, 0.009]

0.553

Combined

 

0.973

0.006 [0.003, 0.010]

 < 0.001*

  1. *Statistically significant after Bonferroni correction with adjusted critical p-value of 0.006