An artificial intelligence-powered PD-L1 combined positive score (CPS) analyser in urothelial carcinoma alleviating interobserver and intersite variability

Kyu Sang Lee ¹, Euno Choi ², Soo Ick Cho ³

⁰Department of Pathology, Seoul National University Bundang Hospital, Seoul National University College of Medicine, Seongnam-si, Republic of Korea.

Histopathology +

|

March 13, 2024

View abstract on PubMed

Summary

Artificial intelligence (AI) improves pathologist agreement in quantifying PD-L1 expression for urothelial carcinoma. This AI tool enhances consistency in PD-L1 combined positive score (CPS) analysis, especially in multi-institutional settings.

Area Of Science

Oncology
Pathology
Artificial Intelligence

Background

Immune checkpoint inhibitors targeting PD-L1 show promise in urothelial carcinoma (UC).
The combined positive score (CPS) quantifies PD-L1 22C3 expression but faces inter-pathologist variability.
Variability stems from assessing both immune and tumor cell positivity.

Purpose Of The Study

To develop and validate an AI-powered tool for quantifying PD-L1 expression in UC.
To assess the AI tool's ability to improve consistency in PD-L1 combined positive score (CPS) assessment.
To evaluate the AI tool's impact on inter-pathologist agreement, particularly in multi-institutional contexts.

Main Methods

Developed an AI PD-L1 CPS analyzer using 1,275,907 cells and 6175.42 mm² of pathologist-annotated tissue from 400 UC whole slide images.
Validated the AI model on 543 UC PD-L1 22C3 cases from three institutions.
Pathologists re-evaluated discrepancy cases with AI assistance, and agreement rates were compared.

Main Results

The AI model achieved 89.5% agreement with the consensus of two or more pathologists.
Pathologist agreement increased from 82.1% to 93.9% after using the AI as a guide.
The AI tool mitigated discordance related to hospital source, specimen type, T stage, histology, and PD-L1 cell type.

Conclusions

AI models can significantly reduce inter-pathologist discrepancies in quantifying PD-L1 22C3 CPS.
AI tools are particularly valuable for improving consistency in telepathology and multi-institutional studies.
This AI solution enhances the reliability of PD-L1 biomarker assessment in urothelial carcinoma.

Keywords:

artificial intelligence combined positive score deep learning programmed death‐ligand 1 urothelial carcinoma