PD-L1 expression assessment in Angiosarcoma improves with artificial intelligence support.

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence

Background:

• Tumoral PD-L1 expression is crucial for immunotherapy selection in cancer treatment.
• Accurate assessment of tumor proportion score (TPS) is challenging for pathologists due to time constraints, leading to low inter-observer concordance.
• Rare cancer types pose a data bottleneck for developing AI-based predictive tools.

Purpose of the Study:

• To develop and validate an AI-powered pipeline for predicting tumor proportion score (TPS) in cancer.
• To address the challenge of limited training data for rare cancers using pre-trained, generalist models.
• To evaluate the practical utility of AI-assisted TPS scoring in improving pathological assessments.

Main Methods:

• Developed and open-sourced a computational pipeline leveraging pre-trained models for TPS prediction.
• Applied the pipeline to predict TPS, particularly for rare cancer types like angiosarcoma.
• Facilitated a feedback loop where pathologists reassessed cases with significant AI-pathologist disagreement.

Main Results:

• The AI pipeline demonstrated strong TPS prediction performance even with limited training data.
• Pathologists frequently updated their TPS scores after reviewing AI predictions that differed significantly from their initial assessment.
• The AI tool served as a valuable 'second opinion,' enhancing diagnostic accuracy.

Conclusions:

• AI-based TPS prediction tools are technically feasible and practically valuable for assisting pathologists, especially in rare cancers.
• Leveraging generalist, pre-trained models overcomes data limitations for developing AI diagnostic aids in underrepresented cancer types.
• AI assistance has the potential to improve decision-making quality in cancer immunotherapy selection.