Comparing deep learning and pathologist quantification of cell-level PD-L1 expression in non-small cell lung cancer whole-slide images

Leander van Eekelen ¹, Joey Spronck ², Monika Looijen-Salamon ²

⁰Department of Pathology, Radboud University Medical Center, P.O.Box 9101, 6500 HB, Nijmegen, The Netherlands. leander.vaneekelen@radboudumc.nl.

Scientific Reports +

|

March 27, 2024

View abstract on PubMed

Summary

A new deep learning algorithm detects programmed death-ligand 1 (PD-L1) expression at the cellular level. This AI shows promise for improving immune-checkpoint inhibitor eligibility assessment by quantifying tumor proportion score (TPS).

Area Of Science

Computational pathology
Artificial intelligence in oncology
Immunohistochemistry analysis

Background

Programmed death-ligand 1 (PD-L1) expression, assessed by tumor proportion score (TPS), is crucial for immune-checkpoint inhibitor eligibility.
Current clinical assessment of PD-L1 TPS suffers from high interobserver variability, limiting treatment efficacy.
Existing automated quantification systems often focus on slide-level TPS and lack cellular-level PD-L1 expression analysis.

Purpose Of The Study

To develop and evaluate a deep learning algorithm for cell-level detection of PD-L1 negative and positive tumor cells.
To establish a cell-level reference standard for benchmarking computer vision algorithms in PD-L1 assessment.
To evaluate the algorithm's performance in slide-level tumor proportion score (TPS) quantification against pathologist agreement.

Main Methods

A deep learning algorithm was developed for cell-level PD-L1 expression detection.
The algorithm was evaluated on a multi-centric dataset with a cell-level reference standard created by six expert readers.
Slide-level TPS quantification was assessed by comparing algorithm predictions with those of six pathologists.

Main Results

The algorithm achieved a mean reader-AI F1 score of 0.55, slightly below the interobserver agreement (mean reader-reader F1 score = 0.68), particularly on external clinical data.
Despite cell-level challenges, the AI demonstrated good agreement with pathologists in quantifying TPS (mean reader-AI Cohen's kappa = 0.49), comparable to interobserver agreement (mean reader-reader kappa = 0.54).
The study established a novel benchmark for computer vision algorithms using a multi-reader, multi-assay cell-level PD-L1 dataset.

Conclusions

The developed deep learning algorithm shows potential for accurate cell-level PD-L1 expression detection.
The AI exhibits favorable agreement with pathologists in TPS quantification, offering a promising tool to reduce interobserver variability.
The study's models are publicly released on the Grand-Challenge platform to facilitate further research and development.