Programmed death ligand-1 (PD-L1) immunohistochemistry is used to predict response to anti-PD-L1 therapy. However, intra- and inter-observer variability influence the prediction of PD-L1 expression. Unlike evaluations of non-small cell lung cancer according to PD-L1 expression determined by tumor proportion score, gastric cancer is evaluated using combined positive scores. We aimed to compare the results of digital image analysis and pathologists' interpretations for predicting response to pembrolizumab in gastric cancer patients. Gastric cancer tissues from patients enrolled in a clinical trial of pembrolizumab were reviewed, and 39 cases were analyzed. PD-L1 22C3 PharmDx (Dako) immunohistochemistry slides were interpreted by digital image analysis and by the consensus of two pathologists, and the results of interpretations were compared with the eventual clinical responses to pembrolizumab. In direct comparisons of digital image analyses and pathologists' interpretations, 33 (84.6%) out of 39 cases showed concordant results. In 6 (15.4%) cases, weak staining of PD-L1, anthracotic pigment deposits, or decalcification artifacts caused discordant results, and all the cases were challenging. In statistical analyses, PD-L1 expression as interpreted by pathologists and digital image analysis did not differ significantly for predicting responses to pembrolizumab (P = 0.1856). Digital image analyses and pathologists' interpretations showed concordant results for PD-L1 combined positive scores in most cases, and the prediction performances of each were not significantly different.
Keywords: Digital image analysis; Gastric cancer; PD-L1; Pembrolizumab.