Fig. 2. Tumor/normal classification using CNNs.

a Numbers of tumor and normal slides in test and training sets. b Per-tile classification metrics. Tile-level test set sizes are provided in Supplementary Data 3. c Fraction of tiles predicted to be tumor within each slide. d Per-slide AUC values for tumor/normal classification for ROC and precision-recall curve (PR). e Pearson correlation coefficients between predicted and pathologist evaluation of tumor purity. P values are based on the permutation test of the dependent variable after Bonferroni correction across all cancer types. Raw and adjusted P values are provided in Supplementary Data 4.