|
model / data
|
Cam_ori
|
Cam_HE
|
Cam_TL
|
TL_ori
|
TL_HE
|
TL_Cam
|
|---|
|
ResNet_ori
|
0.85
|
0.00
|
0.12
|
0.00
|
0.00
|
0.54
|
|
ResNet_HE
|
0.02
|
0.79
|
0.28
|
0.13
|
0.55
|
0.01
|
|
ResNet_TL
|
0.36
|
0.10
|
0.79
|
0.55
|
0.08
|
0.24
|
- The training images (from the Camelyon16 data set) are 1) original (Cam_ori), 2) normalized by the CycleGAN to the HEV data set (Cam_HE) or 3) the TL data set (Cam_TL), respectively. For each training sets, a ResNet model was trained: 1) ResNet_ori, 2) ResNet_HE and 3) ResNet_TL. All models were tested on images from the Camelyon16 data set (n=1728 images) and the TL data set (n=1802 images). There were again three versions of both test data sets: one original version (Cam_ori and TL_ori), one version normalized to the HEV data set (Cam_HE and TL_HE), and one version normalized to the Camelyon16 (TL_Cam) or the TL data set (Cam_TL). The best kappa value obtained for each test set (column-wise) on all models is shown in bold
