Table 3. Diagnostic performance of 48 DLR models for low nuclear grade, ER+, PR+, and HER2+ classification.
| Tasks | Methods | DLR models | |||||
|---|---|---|---|---|---|---|---|
| Classifier | ACC | AUC (95% CI) | Sensitivity | Specificity | F1 | ||
| Nuclear grade | ResNet50 | LR | 0.758 | 0.573 (0.521–0.625) | 0.867 | 0.267 | 0.854 |
| SVM | 0.685 | 0.568 (0.522–0.670) | 0.763 | 0.333 | 0.798 | ||
| RF | 0.673 | 0.596 (0.573–0.729) | 0.704 | 0.533 | 0.779 | ||
| XGBoost | 0.818 | 0.633 (0.576–0.749) | 0.919 | 0.367 | 0.892 | ||
| InceptionV3 | LR | 0.806 | 0.509 (0.472–0.647) | 0.947 | 0.100 | 0.890 | |
| SVM | 0.818 | 0.524 (0.380–0.624) | 1 | 0.030 | 0.903 | ||
| RF | 0.812 | 0.562 (0.499–0.689) | 0.948 | 0.267 | 0.898 | ||
| XGBoost | 0.655 | 0.544 (0.511–0.658) | 0.696 | 0.433 | 0.764 | ||
| DenseNet121 | LR | 0.711 | 0.535 (0.416–0.661) | 0.787 | 0.333 | 0.819 | |
| SVM | 0.717 | 0.501 (0.438–0.616) | 0.860 | 0.267 | 0.857 | ||
| RF | 0.778 | 0.562 (0.434–0.676) | 0.887 | 0.233 | 0.869 | ||
| XGBoost | 0.721 | 0.553 (0.463–0.716) | 0.830 | 0.333 | 0.839 | ||
| ER | ResNet50 | LR | 0.647 | 0.592 (0.536–0.689) | 0.806 | 0.340 | 0.751 |
| SVM | 0.680 | 0.531 (0.514–0.606) | 0.990 | 0.075 | 0.803 | ||
| RF | 0.692 | 0.588 (0.569–0.624) | 0.874 | 0.340 | 0.790 | ||
| XGBoost | 0.660 | 0.601 (0.51–0.668) | 0.835 | 0.415 | 0.782 | ||
| InceptionV3 | LR | 0.679 | 0.584 (0.552–0.626) | 0.913 | 0.226 | 0.790 | |
| SVM | 0.660 | 0.515 (0.453–0.650) | 0.864 | 0.264 | 0.771 | ||
| RF | 0.667 | 0.618 (0.592–0.647) | 0.796 | 0.415 | 0.759 | ||
| XGBoost | 0.654 | 0.528 (0.495–0.581) | 0.806 | 0.359 | 0.755 | ||
| DenseNet121 | LR | 0.667 | 0.514 (0.438–0.595) | 0.893 | 0.226 | 0.780 | |
| SVM | 0.660 | 0.606 (0.546–0.692) | 0.669 | 0.642 | 0.723 | ||
| RF | 0.660 | 0.541 (0.468–0.634) | 0.815 | 0.358 | 0.760 | ||
| XGBoost | 0.641 | 0.570 (0.517–0.675) | 0.738 | 0.453 | 0.731 | ||
| PR | ResNet50 | LR | 0.592 | 0.553 (0.437–0.628) | 0.848 | 0.257 | 0.702 |
| SVM | 0.629 | 0.509 (0.446–0.664) | 0.908 | 0.143 | 0.757 | ||
| RF | 0.636 | 0.576 (0.524–0.652) | 0.704 | 0.518 | 0.711 | ||
| XGBoost | 0.662 | 0.675 (0.582–0.779) | 0.694 | 0.554 | 0.712 | ||
| InceptionV3 | LR | 0.632 | 0.542 (0.497–0.640) | 0.837 | 0.364 | 0.720 | |
| SVM | 0.711 | 0.727 (0.673–0.761) | 0.982 | 0.311 | 0.803 | ||
| RF | 0.685 | 0.696 (0.641–0.763) | 0.800 | 0.516 | 0.752 | ||
| XGBoost | 0.696 | 0.755 (0.707–0.806) | 0.755 | 0.608 | 0.748 | ||
| DenseNet121 | LR | 0.566 | 0.504 (0.451–0.553) | 0.767 | 0.303 | 0.667 | |
| SVM | 0.610 | 0.603 (0.581–0.636) | 0.735 | 0.393 | 0.706 | ||
| RF | 0.623 | 0.601 (0.535–0.629) | 0.674 | 0.554 | 0.698 | ||
| XGBoost | 0.649 | 0.592 (0.517–0.724) | 0.704 | 0.554 | 0.718 | ||
| HER2 | ResNet50 | LR | 0.641 | 0.713 (0.656–0.774) | 0.764 | 0.572 | 0.604 |
| SVM | 0.588 | 0.628 (0.581–0.722) | 0.646 | 0.560 | 0.504 | ||
| RF | 0.634 | 0.626 (0.564–0.698) | 0.600 | 0.653 | 0.541 | ||
| XGBoost | 0.635 | 0.597 (0.572–0.645) | 0.563 | 0.670 | 0.500 | ||
| InceptionV3 | LR | 0.608 | 0.506 (0.408–0.552) | 0.309 | 0.776 | 0.362 | |
| SVM | 0.614 | 0.5352 (0.415–0.543) | 0.182 | 0.857 | 0.253 | ||
| RF | 0.680 | 0.582 (0.505–0.631) | 0.400 | 0.837 | 0.473 | ||
| XGBoost | 0.640 | 0.547 (0.489–0.616) | 0.291 | 0.837 | 0.368 | ||
| DenseNet121 | LR | 0.673 | 0.581 (0.568–0.606) | 0.182 | 0.948 | 0.285 | |
| SVM | 0.607 | 0.568 (0.492–0.619) | 0.527 | 0.653 | 0.492 | ||
| RF | 0.634 | 0.634 (0.593–0.701) | 0.673 | 0.612 | 0.569 | ||
| XGBoost | 0.615 | 0.539 (0.493–0.639) | 0.542 | 0.650 | 0.477 | ||
DLR, deep learning radiomics; ER, estrogen receptor; PR, progesterone receptor; HER2, human epidermal growth factor receptor 2; ACC, accuracy; AUC, area under the curve; CI, confidence interval; LR, logistic regression; SVM, support vector machine; RF, random forest; XGBoost, eXtreme Gradient Boosting.