1244P - AI-based accurate PD-L1 IHC assessment in non-small cell lung cancer across multiple sites and scanners

Background

Immunohistochemical PD-L1 expression, a tissue-based biomarker that predicts response to immune checkpoint therapy, is used to guide treatment decisions in advanced non-small cell lung cancer (NSCLC). Artificial intelligence (AI) technology may optimize PD-L1 scoring in regard to standardization, accuracy and efficiency. However, previous such approaches failed to show consistency across samples derived from different sites and scanning devices.

Methods

We developed an AI software for automated PD-L1 tumor proportion scoring (TPS) according to international guidelines. Four pathologists assessed PD-L1 expression on 146 whole-slide images (WSI) of NSCLC biopsies obtained from three institutions, four scanning hardware types and the Ventana SP263 assay. Pathologists selected one representative region of interest (ROI) per WSI and scored PD-L1 without AI assistance (path-only). After a 2-week washout period, the same pathologists were presented with the same ROIs together with AI results (AI-only). Being able to adjust the AI results, pathologists concluded the final scores (AI+path). In addition, TPS scores were determined globally on the WSIs by the AI without human intervention and were compared against clinical WSI scores.

Results

For the threshold of TPS≥1%, inter-rater agreements between a) pathologist and AI+path, b) pathologist and AI-only, and c) AI-only and AI+path were 86.3%, 85.6%, and 89.7%, respectively (for TPS≥50%: a) 85.6%, b) 80.8%, c) 92.5%). With AI assistance, pathologists scored faster compared to manual scoring (median time: 72 vs. 117 sec/ROI; p<0.01). Additionally, scoring of WSIs by the AI without human intervention showed higher agreement with clinical scores than inter-human agreement (<85%) reported in literature.

Conclusions

Using challenging validation data from three institutions and four scanners, ROI scoring with the support of our AI PD-L1 quantifier showed high agreement with pathologists, while reducing assessment time. Of note, fully automatic global AI scoring on WSIs resulted in equally high agreement rates. These results demonstrate suitability and safety of the AI system for application in clinical routine, ultimately improving accuracy and efficiency of PD-L1 scoring.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Mindpeak GmbH.

Funding

Mindpeak GmbH.

Disclosure

R. Erber, C. Schaaf, N. Abele: Financial Interests, Personal, Speaker, Consultant, Advisor: Mindpeak GmbH. R. Banisch, M. Päpper, P. Frey, K. Daifalla, S. Günther: Financial Interests, Personal, Full or part-time Employment: Mindpeak GmbH. A. Hartmann: Financial Interests, Institutional, Research Funding: Mindpeak GmbH. T. Lang: Financial Interests, Personal, Stocks or ownership: Mindpeak GmbH. All other authors have declared no conflicts of interest.