Researchers have developed an artificial intelligence (AI) tool that helps doctors more accurately measure a key marker used to decide which lung cancer patients might benefit from certain treatments. The AI system reduces human variation in interpreting test results and highlights borderline cases, making the overall process more consistent and reliable.
The team’s study centred on improving the accuracy and consistency of PD-L1 Tumour Proportion Score (TPS) assessment – an essential metric used to determine whether lung cancer patients are likely to benefit from immunotherapy.
This research, published in the journal Scientific Reports, is a clear example of how AI can be used to reshape some of the most important diagnostic tests in modern oncology. Scientists at The Institute of Cancer Research, London, and Queen’s University Belfast (QUB) led the study, which was funded by the National Institute for Health and Care Research and carried out in QUB laboratories.
Why PD‑L1 matters in lung cancer treatment
PD‑L1, a protein expressed on the surface of some tumour cells, plays a powerful part in immune evasion. When it binds to its receptor, PD‑1, on T‑cells, it can suppress the immune response and allow cancer to grow unchecked. Immunotherapy drugs that block this interaction have transformed treatment for many people with non‑small cell lung carcinoma (NSCLC) – the most common type of lung cancer.
However, predicting treatment response is dependent on accurate measurement of PD‑L1 levels. Clinicians use PD‑L1 TPS, which represents the percentage of tumour cells expressing PD‑L1, to guide treatment decisions. Established clinical thresholds – less than 1 per cent and 50 per cent or greater – help determine which therapies may be effective. But assessing TPS is difficult even for experienced pathologists, and previous studies have shown significant variability among experts reviewing the same tissue samples.
This inconsistency has real consequences, as the difference between – for example – a TPS of 45 per cent and a TPS of 55 per cent could change a patient's treatment plan.
Traditional PD‑L1 scoring relies on pathologists manually reviewing immunohistochemistry slides. It is a demanding, subjective process influenced by staining quality, sample preparation and visual judgement. The study highlights substantial inter‑ and intra‑observer variation, especially around the clinically significant thresholds. Previous work has shown that even well‑trained specialists disagree more often than expected, despite improvement with standardised training.
Addressing the challenge
Researchers have long sought ways to introduce objective, reproducible technologies into pathology workflows. Digital Image Analysis, including earlier machine‑learning tools, has shown promise. But this new work introduces something more advanced: a deep‑learning system specifically designed to standardise and support PD‑L1 TPS scoring before regulatory approval pathways come into play.
The researchers developed a deep‑learning tool trained to assist with PD‑L1 TPS calculation. They used a validated multiplex immunofluorescence panel to establish “ground truth” values for the algorithm. This provided a robust baseline against which both human and AI evaluations could be compared.
To build this AI tool, the researchers used training data that included precisely annotated tumour cells. They validated the algorithm across a variety of PD‑L1 expression ranges, comparing its performance with expert‑derived reference standards. Importantly, the multiplex immunofluorescence method used for ground truth helped eliminate potential stain‑based inconsistencies.
The system focuses on the two clinically important thresholds around 1 per cent and 50 per cent, highlighting borderline or ambiguous cases for additional manual review by pathologists. It is not intended to replace human interpreters but rather to flag cases where TPS sits near decision‑making boundaries, provide consistent quantification for mid‑range scores and reduce the likelihood of misclassification.
This approach should help pathologists focus their attention where human expertise is most needed.
A clear path to greater consistency
The study’s most significant finding is that the AI tool can reliably identify cases near key decision thresholds, enabling a standardised approach to what is currently a highly variable assessment. By offering pathologists consistent, AI‑assisted quantification, the system may reduce false negatives at low expression levels, decrease over‑classification near the high‑expression threshold and support more reproducible reporting across institutions.
The tool is designed for pre‑regulatory validation, meaning that although it represents an important step towards future clinical integration, it is not yet approved for clinical practice.
However, in the future, an AI‑supported workflow could be especially valuable in resource‑limited settings or high‑volume clinical centres, where pathologist workloads are heavy and experience levels may vary.
More importantly, for people living with lung cancer, more accurate PD‑L1 scoring could directly influence treatment decisions and outcomes. Immunotherapies such as pembrolizumab and nivolumab rely on PD‑L1 expression as a key eligibility marker. Misclassification – either overestimating or underestimating TPS – carries real risks, from unnecessary side effects to missed therapeutic opportunities. By introducing automated consistency checks and a standardised scoring framework, AI may help ensure that patients receive therapies best suited to their tumour biology.
A glimpse into the future of digital pathology
The paper fits into a growing international effort to validate digital and AI‑enhanced pathology, as seen in earlier large‑scale studies comparing AI PD‑L1 analysers with pathologists across thousands of samples. These complementary studies highlight the global medical community’s drive to bring greater precision to immunotherapy decision‑making.
This study is part of a wider wave of innovation in digital pathology, where AI is increasingly being applied to tasks such as tumour segmentation, biomarker quantification and predictive modelling. The implications extend well beyond PD‑L1 scoring.
“A meaningful advance”
Lead author Professor Manuel Salto-Tellez, Group Leader of the Integrated Pathology Group at The Institute of Cancer Research (ICR) and Chair of Molecular Pathology at QUB, said:
“PD‑L1 scoring is one of the cornerstones of immuno‑oncology, but it remains vulnerable to human variability, particularly around the thresholds that matter most for patients. With demonstrated potential to reduce variability and reinforce clinical decision‑making, the deep‑learning system we’ve developed represents a meaningful advance in digital pathology.
“Our goal is not to fully automate PD‑L1 evaluation, but to give pathologists and clinicians a tool that enhances confidence and consistency, particularly in borderline cases. And this study presents a compelling case for integrating AI into one of the most consequential diagnostic pathways in modern oncology.
“If future regulatory pathways are successful, AI‑driven PD‑L1 scoring could soon become a routine part of cancer diagnostics – supporting pathologists, improving reliability and ultimately helping clinicians offer more precise, personalised care to people living with lung cancer.”
Image credit: Gordon Johnson from Pixabay