Figure A9.

(a) ROC curves for conventional classifiers (level-1) used with the DL output probabilities from the ResNet50 (level-0) model and the hand-crafted features (cascade generalization); (b) Figure showing ROC curves only for the ResNet50 (level-0) model, the ensemble of conventional classifiers, and the best conventional level-1 classifier using the ResNet50 probability outputs. The AUC and the AUC for FPR higher than 0.4 are both presented for comparison. All values are rounded to three significant digits.