Artificial Intelligence (AI) for Programmed Death Ligand-1 (PD-L1) Immunohistochemical Assessment in Urothelial Carcinomas: “Teaching” Cell Differentiation to AI Systems

Assessment of Programmed Death-Ligand 1 (PD-L1) immunohistochemical (IHC) expression on tumor cells (TCs) and immune cells (ICs) in bladder cancer (BC) is challenging. Artificial Intelligence (AI) has potential for accurate PD-L1 IHC scoring, but its efficiency remains debatable. Our aim was to compare two AI protocols provided by the free QuPath software (v0.5.1) (Selected Area Interpretation (AI-SAI) and Whole Slide Imaging (AI-WSI)) with manual PD-L1 IHC scoring. A total of 43 BCs were included. PD-L1 IHC was performed using the SP263 clone. The IHC slides were digitized and further imported into QuPath. The PD-L1 positivity threshold was set at 25%. Statistically significant correlations were observed between AI-SAI and manual interpretation for both TCs (r = 0.85) and ICs (r = 0.57). AI-WSI yielded comparable results, with correlation coefficients of r = 0.82 for TCs and r = 0.56 for ICs. However, AI-SAI demonstrated stronger agreement with manual assessment (κ = 0.86) compared to AI-WSI (κ = 0.65). Receiver Operating Characteristic (ROC) analysis further supported the superiority of AI-SAI, with higher AUC values for both TCs (0.96 vs. 0.92) and ICs (0.92 vs. 0.90). Our findings indicate that AI-SAI is preferable to AI-WSI, particularly in BC cases with high PD-L1-positive TC content. Nevertheless, supervision by an experienced pathologist is mandatory.