2022 World Conference on Lung Cancer (ePosters)-EP11.02-002. A Comparative Study of PD-L1 Scoring: Humans versus AI

Pdf Summary

A comparative study was conducted to evaluate the scoring of PD-L1 expression in non-small cell lung cancer (NSCLC) by pathologists and artificial intelligence (AI). PD-L1 expression is an important factor in determining the efficacy of immunotherapy for NSCLC. The study compared the manual scoring of PD-L1 expression by six pathologists with the results obtained from two different AI algorithms.<br /><br />The pathologists independently scored 51 NSCLC cases using a staining technique called SP263. The same set of whole slide images (WSI) was also analyzed using two commercially available PD-L1 algorithms. Algorithm 1 was certified for usage with the SP263 assay, while algorithm 2 was for research use only.<br /><br />The results showed excellent intraobserver agreement among the pathologists. Interobserver agreement between the pathologists was also analyzed and found to be high, with Fleiss' kappa values ranging from 0.488 to 0.803, depending on the tumor proportion score (TPS) cutoff used.<br /><br />When comparing the results of the AI algorithms with the median manual scores of the pathologists, lower agreement was observed. Algorithm 1, which was certified for usage with the assay, showed a Fleiss' kappa value of 0.354, while algorithm 2 had a value of 0.672. These results indicate that there is room for improvement in the performance of the AI algorithms.<br /><br />In conclusion, the study found excellent interobserver agreement among the pathologists in scoring PD-L1 expression in NSCLC. However, the AI algorithms showed lower agreement with the manual scores. Further analysis is needed to understand the discrepancies between manual and AI analysis and to improve the performance of the algorithms.

Asset Subtitle

Luka Brcic

Meta Tag

Speaker Luka Brcic

Topic Pathology - New Technology

Keywords

comparative study

PD-L1 expression

NSCLC

pathologists

artificial intelligence

AI

SP263

whole slide images

interobserver agreement

Fleiss' kappa values