Table 4.
Metrics for discriminative localization of benign and malignant breast masses in the weakly-supervised deep learning algorithm.
| Internal validation | External validation | |||||
|---|---|---|---|---|---|---|
| Benign | Malignant | P value | Benign | Malignant | P value | |
| VGG16 | 1.00 | 1.00 | ||||
| Correct | 99 | 100 | 99 | 100 | ||
| Incorrect | 1 | 0 | 1 | 0 | ||
| ResNet34 | 1.00 | 0.68 | ||||
| Correct | 99 | 100 | 96 | 98 | ||
| Incorrect | 1 | 0 | 4 | 2 | ||
| GoogLeNet | N/A | 0.25 | ||||
| Correct | 100 | 100 | 97 | 100 | ||
| Incorrect | 0 | 0 | 3 | 0 | ||
N/A = not applicable. For discriminative localization, we created binary images by applying a threshold of 0.3 to a class activation map (CAM) and compared them with manual annotation. Discriminative localization is regarded as correct when the segmented area overlaps with the manually annotated area.