Table 8 Classification results.
Learning method | CNN model | Sensitivity | Specificity | Accuracy | Balanced accuracy |
|---|---|---|---|---|---|
Weakly supervised learning | AD MIL LeNet-like | 0.893 | 0.907 | 0.898 | 0.900 |
Conventional MIL pooling LeNet-like | 0.921 | 0.778 | 0.873 | 0.850 | |
AD MIL AlexNet-like | 0.930 | 0.889 | 0.916 | 0.910 | |
Conventional MIL pooling AlexNet-like | 0.893 | 0.750 | 0.845 | 0.822 | |
AD MIL Inception | 0.874 | 0.880 | 0.876 | 0.877 | |
Conventional MIL pooling Inception | 0.897 | 0.528 | 0.773 | 0.713 | |
AD MIL ResNet | 0.874 | 0.917 | 0.888 | 0.900 | |
Conventional MIL pooling ResNet | 0.921 | 0.778 | 0.873 | 0.850 | |
AD MIL DenseNet | 0.822 | 0.315 | 0.652 | 0.569 | |
Conventional MIL pooling DenseNet | 1.000 | 0.000 | 0.665 | 0.500 | |
Supervised learning: image-based evaluation | AlexNet-like | 0.898 | 0.848 | 0.880 | 0.873 |
Supervised learning: case-based evaluation | AlexNet-like | 0.985 | 0.713 | 0.849 | 0.849 |
