Impact of artificial intelligence assistance on combined PD-L1 scoring on whole immunohistochemical slide image of tumors.

Abstract

e16006 Background: Immune checkpoint blockade targeting PD-L1 has transformed cancer treatment, making accurate evaluation of PD-L1 expression critical for immunotherapy eligibility. In practice, PD-L1 assessment, particularly using the Combined Positive Score (CPS) is challenged by interpretative variability among pathologists. Artificial intelligence (AI) based approaches may help address these limitations by enabling more consistent and exhaustive PD-L1 quantification instead of visual estimation by pathologists. Here, we assess the impact of the DiaKwant PD-L1 CPS, an AI solution, on the scoring of pathologists, evaluated against an expert-consensus reference across multiple tumor types. Methods: A total of 149 immunostained-tissue samples from gastrointestinal and head and neck tumors were digitized and independently scored by three expert pathologists. Discordant cases were jointly reviewed to establish a consensus reference standard. Four different pathologists from various institutions independently reviewed all cases without AI support. Following a washout period of two months, the pathologists reassessed the cases with AI assistance. For each slide, the algorithm provided CPS values along with a PD-L1 expression heatmap to support result interpretability. Pathologist assessments were compared with the reference standard to quantify the impact of AI on scoring accuracy. Performance was assessed using accuracy, sensitivity and specificity for clinically relevant binary cutoffs, and intraclass correlation coefficient (ICC) to evaluate concordance between pathologists. Scoring time was also recorded. Results: Using organ-specific cut-offs, the CPS accuracy increased from 77.6% to 83.9% with AI assistance (p < 0.01), sensitivity increased from 78.5% to 89.6% and specificity increased from 73.4% to 76.2%, while the average scoring time per case decreased from 109s to 78s (p < 0.01). The ICC between pathologists rose from 0.71 to 0.90 when introducing AI assistance. Conclusions: AI-assisted PD-L1 assessment significantly improved agreement with the reference standard, enhanced inter-pathologist reliability, and reduced scoring time while maintaining clinically relevant performance across tumor types. These findings support the potential of AI assistance to exhaustively quantify PD-L1 CPS and facilitate more standardized and robust patient selection for immunotherapy in routine practice.