Fig. 2. Convolutional neural networks achieve high performance in the prediction of PD-L1 and PD-1 expression.

Receiver operating characteristics (ROC) curves for the performance of the proposed models, in terms of AUC, for PD-L1 and PD-1 prediction in the BCCA and MA31 cohorts. a The model obtained high prediction accuracies for both the BCCA cross-validation (0.911) and BCCA test set (0.915). When analyzing only concordant cases between pathologists, AUC performance was further increased (0.928). b For the external MA31 cohort, the performance dropped to 0.854, showing that a calibration step may benefit the application of the system to new cohorts. Indeed, the calibration step increased the AUC on MA31 to 0.886, which was further increased to 0.919 after removing the discordant cases. c The AUC performance results for PD-1 prediction were lower than for PD-L1. The PD-1 AUC results were high, however, given the extremely imbalanced nature of data (only 3% positives), which poses optimization difficulties due to very few positive samples to train the system with.