Comparative performance of PD-L1 scoring by pathologists and AI algorithms.

Area of Science:

• Oncology
• Pathology
• Biomarker Analysis

Background:

• Immune-checkpoint inhibitors have transformed non-small cell lung carcinoma (NSCLC) treatment.
• Tumour proportion score (TPS) of PD-L1 expression is a key biomarker for predicting treatment response.
• Accurate PD-L1 scoring is crucial for guiding therapeutic decisions in NSCLC.

Purpose of the Study:

• To compare the effectiveness of pathologists and artificial intelligence (AI) algorithms in scoring PD-L1 expression in NSCLC.
• To assess the agreement among pathologists and between AI tools and pathologists.
• To evaluate the reliability of AI in PD-L1 TPS assessment for clinical application.

Main Methods:

• Scoring of 51 SP263-stained NSCLC cases by six pathologists using light microscopy and whole-slide images (WSI).
• Evaluation of cases using two commercial AI software tools: uPath and Visiopharm.
• Analysis of intra- and interobserver agreement among pathologists and comparison of AI scores with median pathologist scores at TPS cutoffs of 1% and 50%.

Main Results:

• Pathologists demonstrated moderate interobserver agreement for TPS <1% (Fleiss' kappa 0.558) and almost perfect agreement for TPS ≥50% (Fleiss' kappa 0.873).
• High intraobserver consistency was observed among pathologists (Cohen's kappa 0.726–1.0).
• AI algorithms showed fair agreement (uPath, Fleiss' kappa 0.354) and substantial agreement (Visiopharm, Fleiss' kappa 0.672) with median pathologist scores at the 50% TPS cutoff.

Conclusions:

• Pathologists exhibit strong concordance in PD-L1 TPS scoring, particularly at higher expression levels.
• Current AI algorithms show less consistent performance compared to expert human evaluation.
• Further AI tool development is necessary to achieve the reliability required for critical clinical decision-making in NSCLC treatment.