A new AI-assisted scoring system for PD-L1 expression in NSCLC

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Programmed cell death ligand-1 (PD-L1) expression is a critical biomarker in non-small cell lung cancer (NSCLC).
• Accurate scoring of PD-L1 expression, often using the tumor proportion score (TPS), is essential for treatment decisions.
• Artificial intelligence (AI) offers a potential tool to standardize and improve PD-L1 scoring.

Purpose of the Study:

• To evaluate a novel AI-assisted scoring system for PD-L1 expression in NSCLC.
• To compare the performance of the Aitrox AI model against experienced and inexperienced pathologists.
• To assess the utility of AI in routine pathological diagnosis of NSCLC.

Main Methods:

• PD-L1 expression was categorized into negative (TPS < 1%), low (1-49%), and high (≥50%) using the tumor proportion score (TPS).
• The Aitrox AI segmentation model was trained, validated, and tested on whole slide images (WSIs).
• AI model performance was compared to TPS readings from experienced pathologists, inexperienced pathologists, and a Gold Standard established by expert review.

Main Results:

• The Aitrox AI model demonstrated strong correlation with the TPS Gold Standard, comparable to experienced pathologists.
• AI performance exceeded that of inexperienced pathologists, particularly in negative and low PD-L1 expression groups.
• The AI model showed limitations in high PD-L1 expression groups and cases with significant false-positive signals.

Conclusions:

• The Aitrox AI model shows potential as an assistive tool for pathologists in scoring PD-L1 expression in NSCLC.
• AI-assisted scoring can enhance diagnostic consistency and efficiency in routine pathology.
• Further refinement is needed for optimal performance across all PD-L1 expression levels.