Back to List

Abstract : Assistance with an artificial intelligence-powered PD-L1 analyzer reduces interobserver variation in pathologic reading of tumor proportion score in non-small cell lung cancer

S. Choi et al. — ESMO 2021



Programmed death ligand 1 (PD-L1) expression is the standard biomarker in advanced non-small cell lung cancer (NSCLC). However, manual evaluation of PD-L1 tumor proportion score (TPS) by pathologists has practical limitations of interobserver bias, variation in subjectivity on the area of interest, and intensive labor. This study aimed to explore whether the artificial intelligence (AI)-powered TPS analyzer could reduce the human discrepancy.


AI-powered TPS analyzer, namely Lunit SCOPE PD-L1, was developed with a total of 393,565 tumor cells annotated by board-certified pathologists for PD-L1 expression in 802 whole-slide images (WSI) of NSCLC stained by 22C3 pharmDx immunohistochemistry. Three independent pathologists labeled PD-L1 TPS into 3 class categories: TPS < 1%, 1-49%, or ≥ 50%, of 479 NSCLC slides. For the cases of disagreement between each pathologist and AI model, the pathologists were asked to revise TPS class in assistance with AI model which not only detects PD-L1 positivity of tumor cells, but also calculates WSI-level TPS. Finally, we compared the concordance rate of three pathologists with or without AI assistance.


Without AI assistance, 3 pathologists concordantly labeled TPS in 81.4% of cases (n = 390 / 479, κ = 0.798), and the concordance rate between the consensus of pathologists and standalone AI model was 86.4% (n = 337 / 390). Afterward, pathologists revised their initial labeling with assistance of AI model for the cases of disagreement between the pathologist and AI model (n = 91, 93, and 107, respectively for each pathologist). Interestingly, the overall concordance rate of three pathologists with AI assistance was increased to 90.2% (n = 432 / 479, κ = 0.890). Subgroup analysis showed that the concordance rates without AI assistance according to PD-L1 TPS <1%, 1-49%, and ≥50% class were 67.9%, 72.2%, and 92.4%, respectively, which were increased with AI assistance to 89.6%, 86.2%, and 93.6%, respectively.


Assistance with AI-powered TPS analyzer substantially improved the pathologist’s consensus and could be regarded as a reference for the final labeling of TPS, especially in the subgroups of TPS <1% and 1-49%.

Read the full paper

S. Choi1, S. Kim2, H. Kim3, S. Cho4, M. Ma4, S. Park4, S. Pereira4, B.J. Aum4, S. Shin4, K. Paeng4, D. Yoo4, W. Jung4, C. Ock4, S. Lee5, Y. Choi1, J. Chung3, T.S. Mok6

1 Department Of Pathology And Translational Genomics, Samsung Medical Center, Sungkyunkwan University School of Medicine, 135-710 - Seoul/KR

2 Department Of Pathology, Ajou University School of Medicine, 16499 - Suwon/KR

3 Department Of Pathology, Seoul National University Bundang Hospital, 13620 - Seongnam/KR

4 Oncology Group, Lunit Inc., 06241 - Seoul/KR

5 Division Of Hematology-oncology, Department Of Medicine, Samsung Medical Center, Sungkyunkwan University School of Medicine, 135-710 - Seoul/KR

6 Department Of Clinical Oncology, State Key Laboratory of Translational Oncology and Chinese University of Hong Kong, 999077 - Shatin/HK

ESMO 2021

Read more