Fig. 3

ROC curves of baseline models, KD and PD for clinical response prediction in test set. The EfficientNetB7 model with PD achieved a high accuracy and substantially improved upon baseline models and KD. Sensitivity and specificity for clinician predictions are reported, both as an average and for the individual clinicians with 3 (C3), 20 (C20), and 40 (C40) years of experience. ROC = receiver operating characteristic, KD = Knowledge Distillation, PD = Progressive Distill.