Abstract
PD-L1 expression is an important biomarker for selecting patients who are eligible for immune checkpoint inhibitor (ICI) therapy. However, evaluating PD-L1 through immunohistochemistry often faces significant interobserver variability and requires considerable time and resources. Recent advancements in artificial intelligence (AI) have transformed the field of pathology, leading to more standardized and reproducible methods for biomarker quantification. In this study, we examine the application of AI-driven models, particularly deep learning algorithms, to predict PD-L1 expression directly from hematoxylin and eosin-stained histological slides. Several AI-based approaches have been studied, demonstrating high accuracy in estimating PD-L1 expression and predicting responses to ICIs across various cancer types. AI-driven assessments of PD-L1 have been shown to reduce the subjectivity associated with manual scoring methods, such as the Tumor Proportion Score and the Combined Positive Score. Moreover, integrating AI with multimodal data, including genomics, radiomics, and real-world clinical data, can further enhance predictive accuracy and improve patient stratification for immunotherapy. Finally, AI-driven computational pathology offers a transformative approach to biomarker evaluation, providing a faster, more objective, and cost-effective alternative to traditional methods, with significant implications for personalized oncology and precision medicine. Despite these promising results, several challenges remain to be addressed, such as the need for large-scale validation, standardization of AI models, and regulatory approvals for clinical implementation. Tackling these issues will be crucial for incorporating AI-based PD-L1 assessments into routine pathology workflows.