AI Matches Pathologist Accuracy in PD-L1 Scoring for NSCLC

Artificial intelligence may offer a more consistent and, in some cases, more accurate way to assess PD-L1 expression in non-slight cell lung cancer (NSCLC) than traditional pathologist review, according to research presented at the European Lung Cancer Congress in Copenhagen in March 2026. This development could refine patient selection for immunotherapy, a treatment increasingly guided by PD-L1 biomarker testing.

PD-L1, or programmed death-ligand 1, is a protein found on cancer cells that helps them evade the immune system. Immunotherapies targeting the PD-1/PD-L1 pathway aim to block this interaction, allowing the immune system to recognize and attack cancer cells. However, accurately measuring PD-L1 levels is crucial for determining which patients are most likely to benefit from these therapies. Current standard practice relies on immunohistochemistry (IHC), a technique where pathologists visually assess the percentage of tumor cells expressing PD-L1. This manual assessment can be subject to variability between different pathologists and even within the same pathologist over time.

Blueprint Project and the Quest for Standardization

The need for standardized PD-L1 assessment emerged nearly a decade ago with the introduction of multiple PD-L1 assays, each linked to a specific immunotherapy drug. “At that time, we were focusing on how to select patients for immunotherapy,” explained Fred R. Hirsch, MD, PhD, professor of medicine and pathology and executive director of the Center of Excellence for Thoracic Oncology at Mount Sinai Tisch Cancer Center, in an interview with Healio. “Each company had their own assay to assess PD-L1 expression… That is not practical or useful for the pathologist.” Healio

The Blueprint Programmed Death Ligand 1 Immunohistochemistry Comparability Project sought to address this issue by evaluating the consistency of different assays. Researchers found that three assays yielded similar results, narrowing the field. However, inherent variability in manual assessment remained a concern. The goal, as Dr. Hirsch stated, is to “be sure that we offer the right treatment to the right patient,” a cornerstone of personalized medicine.

AI Steps In: Non-Inferiority and Potential Improvements

Recent research investigated whether AI could provide a more objective and consistent method for PD-L1 scoring. Researchers analyzed 80 slides from the original Blueprint project, comparing AI-derived scores to those from 24 pathologists. The findings, presented at the European Lung Cancer Congress, demonstrated that AI was “at least as good as manual, and in some cases better” at quantifying PD-L1 expression. Healio

Specifically, the AI and pathologists showed high agreement at a 50% cutoff, a critical threshold for treatment decisions. Current guidelines recommend immunotherapy monotherapy for patients with a tumor proportion score (TPS) of 50% or higher, while those with lower scores typically receive a combination of immunotherapy and chemotherapy. The AI platform demonstrated particularly strong performance with the SP142 assay, showing a significant improvement in PD-L1 scoring compared to manual assessment.

Understanding the Implications for Treatment

The potential benefits of AI-driven PD-L1 scoring extend beyond simply reducing variability. Accurate PD-L1 assessment is vital for appropriate treatment selection. As Dr. Hirsch explained, patients with high PD-L1 expression may not benefit from the added toxicity of chemotherapy, making accurate scoring essential for avoiding unnecessary side effects. Healio

The study highlights the evolving role of AI in pathology. While AI is not intended to replace pathologists, it can serve as a valuable tool to enhance accuracy, efficiency, and consistency in biomarker assessment. This is particularly important in community settings where access to specialized expertise may be limited.

Limitations and Future Directions

While the initial results are promising, it’s important to acknowledge the study’s limitations. The analysis was based on a relatively small sample size of 80 slides from the Blueprint project. Further research is needed to validate these findings in larger, more diverse patient populations. The study focused on comparing AI to pathologist consensus, rather than evaluating its impact on real-world clinical outcomes.

The next step, according to Dr. Hirsch, is to correlate AI-derived PD-L1 scores with treatment response in a large clinical trial. “The next step would be to take the AI platform and apply it to a large clinical trial or a large series of patients and see if prediction is better for benefit of immunotherapy by using the AI platform than manual assessment.” This will help determine whether AI can truly improve patient outcomes and refine the use of immunotherapy in NSCLC.

Broader Context: AI in Cancer Diagnostics

The application of AI to PD-L1 scoring is part of a broader trend toward the use of artificial intelligence in cancer diagnostics. AI algorithms are being developed to assist pathologists in a variety of tasks, including identifying cancerous cells, grading tumors, and predicting treatment response. Healio These tools have the potential to improve the accuracy and efficiency of cancer diagnosis, ultimately leading to better patient care.

The development and validation of AI-powered diagnostic tools require careful attention to data quality, algorithm transparency, and clinical validation. It is crucial to ensure that these tools are accurate, reliable, and equitable across diverse patient populations. Ongoing research and collaboration between pathologists, data scientists, and clinicians will be essential to realize the full potential of AI in cancer care.

For more information:

Fred R. Hirsch, MD, PhD, can be reached at [email protected].