Table 3.
Comparison of average fivefold cross-validation AUROCs in transfer learning models trained with full B-mode images and patch-based images of liver parenchyma using data from 397 training/validation patients. The higher result between full-image and patch inputs for each model is bolded.
| Model | Input B-mode | AUROC |
|---|---|---|
| VGG-16 | Full image | 0.773 ± 0.073 |
| Patches | 0.808 ± 0.061 | |
| ResNet-50 | Full image | 0.762 ± 0.108 |
| Patches | 0.830 ± 0.035 | |
| Inception V3 | Full image | 0.757 ± 0.130 |
| Patches | 0.818 ± 0.048 | |
| DenseNet-121 | Full image | 0.770 ± 0.116 |
| Patches | 0.824 ± 0.032 |