Automated tumor proportion scoring for PD-L1 expression based on multistage ensemble strategy in non-small cell lung cancer

Introduction Programmed cell death ligand-1 (PD-L1) expression is a promising biomarker for identifying treatment related to non-small cell lung cancer (NSCLC). Automated image analysis served as an aided PD-L1 scoring tool for pathologists to reduce inter- and intrareader variability. We developed a novel automated tumor proportion scoring (TPS) algorithm, and evaluated the concordance of this image analysis algorithm with pathologist scores. Methods We included 230 NSCLC samples prepared and stained using the PD-L1(SP263) and PD-L1(22C3) antibodies separately. The scoring algorithm was based on regional segmentation and cellular detection. We used 30 PD-L1(SP263) slides for algorithm training and validation. Results Overall, 192 SP263 samples and 117 22C3 samples were amenable to image analysis scoring. Automated image analysis and pathologist scores were highly concordant [intraclass correlation coefficient (ICC) = 0.873 and 0.737]. Concordances at moderate and high cutoff values were better than at low cutoff values significantly. For SP263 and 22C3, the concordances in squamous cell carcinomas were better than adenocarcinomas (SP263 ICC = 0.884 vs 0.783; 22C3 ICC = 0.782 vs 0.500). In addition, our automated immune cell proportion scoring (IPS) scores achieved high positive correlation with the pathologists TPS scores. Conclusions The novel automated image analysis scoring algorithm permitted quantitative comparison with existing PD-L1 diagnostic assays and demonstrated effectiveness by combining cellular and regional information for image algorithm training. Meanwhile, the fact that concordances vary in different subtypes of NSCLC samples, which should be considered in algorithm development.

Automated summary

A novel automated image analysis algorithm was developed for PD-L1 scoring in NSCLC, showing high concordance with pathologist scores. The algorithm demonstrated effectiveness by combining cellular and regional information, and concordances varied in different subtypes of NSCLC samples.