A Deep Learning-Based Assay for Programmed Death Ligand 1 Immunohistochemistry Scoring in Non-Small Cell Lung...

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Programmed death ligand 1 (PD-L1) expression is crucial for predicting response to immunotherapy in non-small cell lung carcinoma (NSCLC).
• Accurate assessment of PD-L1 tumor proportion score (TPS) by pathologists is essential but can be challenging, leading to inter-observer variability.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) system to assist pathologists in calculating the PD-L1 TPS for the 22C3 assay in NSCLC.
• To assess the impact of AI-assisted systems on the accuracy and consistency of pathologists' TPS assessments.

Main Methods:

• An AI model was developed to calculate PD-L1 TPS, offering automated scoring and visualized overlay images.
• Four AI-assisted assessment methods were evaluated, involving pathologists' initial assessment followed by AI review, or AI review prior to pathologist assessment.
• Mixed-model analysis was used to compare the accuracy of AI-assisted methods against pathologists' independent scoring, particularly in discordant cases.

Main Results:

• Overall, AI-assisted TPS assessment did not significantly alter pathologists' scoring accuracy across all samples.
• In cases with discordant pathologist judgments (332 samples), AI-assisted methods significantly improved assessment accuracy (ORs ranging from 1.28 to 1.29, P < .012).
• No significant differences in accuracy were observed among the four AI-assisted methods when analyzing discordant cases.

Conclusions:

• AI-assisted systems show significant utility in improving the accuracy of PD-L1 TPS assessment in NSCLC, especially in challenging cases with inter-observer variability.
• The developed AI system can serve as a valuable tool to support pathologists, enhancing diagnostic consistency and potentially improving treatment decisions for NSCLC patients.