Fig. 5From: Visualizing histopathologic deep learning classification and anomaly detection using nonlinear feature space dimensionality reductionComposite approaches to WSI classification. a ROC curves generated using a hybrid prediction score comprised of a blend of the percentage of tile distribution and probability scores across the different classes. This approach achieved an even higher AUC. b Relationship of accuracy (y-axis) of this hybrid classifier at different “undefined” score cutoffs of the t-SNE component (% of tiles in undefined space). c Relationship of accuracy of another more conservative scoring approach where only concordant cases are classified at different “undefined” score cutoffs. This approach further reduces the overall error rate with a minimal change to the number of classified cases compared to t-SNE aloneBack to article page