ICR's New AI Pathology Tool Revolutionises Complex Lung Cancer Diagnoses in the UK

Breakthrough in Lung Cancer Diagnostics: ICR and QUB Unveil AI-Powered Pathology Tool

In a significant advancement for oncology, researchers from the Institute of Cancer Research (ICR) in London and Queen's University Belfast (QUB) have developed a deep-learning artificial intelligence tool designed to enhance the accuracy of PD-L1 scoring in non-small cell lung cancer (NSCLC) diagnoses. This innovation addresses one of the most critical challenges in modern pathology: the subjective variability in assessing PD-L1 expression levels, a key biomarker that determines eligibility for life-saving immunotherapies.

NSCLC, the predominant form of lung cancer accounting for approximately 85 percent of all cases, often presents complex histological features that make precise diagnosis demanding. Traditional manual scoring by pathologists, while expert-driven, can lead to inconsistencies, particularly around pivotal clinical thresholds. The new AI system steps in as a supportive ally, standardising measurements and flagging ambiguous cases for closer human scrutiny, potentially transforming how UK clinicians approach treatment decisions.

This collaboration between ICR—a world-leading cancer research institute closely affiliated with the University of London—and QUB's Centre for Cancer Research and Cell Biology underscores the pivotal role of UK higher education institutions in pioneering digital pathology solutions. As lung cancer remains the UK's leading cause of cancer death, with around 66,000 new cases projected annually, such tools promise to refine patient pathways within the National Health Service (NHS).

Understanding PD-L1: The Biomarker at the Heart of Immunotherapy Decisions

Programmed death-ligand 1, commonly abbreviated as PD-L1, is a protein expressed on the surface of tumour cells in NSCLC. It interacts with PD-1 receptors on T-cells, effectively putting the brakes on the immune system's anti-cancer response. Drugs like pembrolizumab (Keytruda) and nivolumab (Opdivo)—checkpoint inhibitors—block this interaction, unleashing T-cells to attack the cancer.

The Tumour Proportion Score (TPS) quantifies PD-L1 expression as the percentage of viable tumour cells showing partial or complete membrane staining. Clinical guidelines, such as those from the National Institute for Health and Care Excellence (NICE), use thresholds like TPS ≥50 percent for first-line pembrolizumab monotherapy or ≥1 percent for combination therapies. Step-by-step, pathologists examine immunohistochemistry (IHC)-stained slides under a microscope, count positive versus total tumour cells, and assign a score. However, factors like staining variability, tumour heterogeneity, and observer fatigue introduce discrepancies.

In the UK context, where immunotherapy has revolutionised advanced NSCLC outcomes—extending median survival from months to years for responders—accurate PD-L1 assessment is non-negotiable. Misclassification at thresholds can deny patients effective treatment or expose them to unnecessary immune-related toxicities like pneumonitis or colitis.

The Challenge of Variability in Manual PD-L1 Scoring

Pathologists are highly trained, yet human interpretation of PD-L1 IHC slides is inherently subjective. Studies highlight inter-observer variability rates exceeding 20 percent around the 1 percent and 50 percent cut-offs, influenced by slide preparation, microscope optics, and individual experience. Intra-observer variation—disagreement with one's own prior scoring—further complicates reliability.

For instance, a TPS hovering at 45-55 percent might be scored as below or above 50 percent by different experts, altering treatment from chemotherapy alone to immunotherapy plus chemo. In high-volume NHS labs, where lung cancer biopsies surge amid rising incidence linked to smoking, air pollution, and ageing populations, this inconsistency strains resources and patient trust.

UK-specific audits reveal similar issues: a Royal College of Pathologists survey noted up to 15 percent discordance in PD-L1 reporting across labs. This underscores the need for objective aids, positioning AI as a timely intervention in the NHS's digital transformation agenda.

How the ICR-QUB AI Tool Works: A Step-by-Step Breakdown

The deep-learning model employs a U-Net architecture with a ResNet-34 backbone, trained on over 1,100 digital whole-slide images (WSIs) of PD-L1-stained NSCLC samples from the Northern Ireland Biobank. Here's the process:

• Digital Scanning: Routine IHC slides (Ventana SP263 assay) and paired H&E stains are scanned at high resolution using an Aperio AT2 scanner.
• Annotation and Training: Pathologists meticulously annotate tumour regions of interest (ROIs) and individual cells as PD-L1-positive/negative or background. Data augmentation (rotations, flips) enhances robustness; training uses weighted cross-entropy loss over 100 epochs.
• Ground Truth Validation: Multiplex immunofluorescence (mIF) panels (PD-L1, cytokeratin for tumour cells, CD68 for macrophages) confirm annotations, excluding non-tumour PD-L1 expression.
• Inference and Assistance: The AI computes TPS = (positive tumour cells / total tumour cells) × 100, correlating 96.97 percent with pathologist scores. It flags 'borderline' cases (e.g., 1-5 percent or 40-60 percent) for manual review, auto-accepts confident mid-range scores.
• Output: Heatmaps highlight uncertain regions, integrating seamlessly into pathology workflows via QuPath software.

This hybrid human-AI approach preserves pathologist oversight while leveraging computational precision.

Key Findings: Superior Consistency and Threshold Precision

Tested on independent validation sets, the tool demonstrated high pixel-level accuracy, sensitivity, and specificity, with object-level precision/recall affirming cell classification fidelity. Notably, it minimised false negatives below 1 percent and over-scoring near 50 percent—common pitfalls in manual reads.

Achieving near-perfect TPS concordance in low/high expression cohorts, the system refines clinical intervals: confident scoring for 5-40 percent, assisted review elsewhere. Funded by the National Institute for Health and Care Research (NIHR), this pre-regulatory validation paves the way for prospective trials. Limitations include dependence on scan quality and need for diverse ethnic datasets, though its NSCLC focus (adenocarcinoma/squamous) ensures broad applicability.

Spotlight on UK Higher Education: ICR and QUB's Collaborative Excellence

ICR, a postgraduate university specialising in cancer research and part of the University of London, excels in translational pathology. Professor Manuel Salto-Tellez, lead author and ICR Group Leader, also holds the Chair of Molecular Pathology at QUB, exemplifying cross-institutional synergy.

QUB's pathology labs hosted the study, leveraging advanced digital infrastructure. This partnership reflects UK higher education's strength in AI-health intersections, supported by funding bodies like UK Research and Innovation (UKRI). For aspiring researchers, such projects highlight opportunities in computational pathology PhDs and fellowships at these institutions.

Quote from Professor Salto-Tellez: “PD-L1 scoring is one of the cornerstones of immuno-oncology, but it remains vulnerable to human variability... Our goal is not to fully automate, but to give pathologists a tool that enhances confidence and consistency.”

Impact on UK Patients: Tailored Treatments and Survival Gains

With lung cancer claiming over 34,000 UK lives yearly—more than breast and prostate combined—precise diagnostics matter. Immunotherapy responders see median progression-free survival double versus chemo alone. Accurate PD-L1 could expand access, reducing overtreatment costs (£10,000+ per cycle) and side effects.

In NHS settings, where waiting times for diagnostics average 40 days, AI could accelerate reporting, aligning with the 28-day faster diagnosis standard. Real-world data shows PD-L1-high patients (≥50 percent) achieve 20-30 percent five-year survival on pembrolizumab, versus 5-10 percent historically.

AI's Growing Role in UK Pathology Research and Education

UK universities lead globally: University College London (UCL) and Imperial College pioneer AI for tumour segmentation; Edinburgh advances predictive models. The NHS Pathology Transformation Programme mandates AI-ready digital labs by 2028, with 50 percent of trusts already scanning slides.

• Benefits: 24/7 analysis, workload reduction (pathologists handle 20,000+ cases/year), remote expertise sharing.
• Risks: Algorithm bias, data privacy (GDPR-compliant), need for validation.
• Training: Pathology curricula now include AI modules at QUB, ICR-linked MSc programmes.

This ICR tool exemplifies scalable innovation, potentially integrable into NHS platforms like those from PathAI or Paige.AI partners.

Challenges, Ethical Considerations, and Regulatory Pathways

Adoption hurdles include scanner standardisation, staff upskilling, and MHRA approval as Class IIb devices. Ethical AI demands diverse training data to mitigate bias against underrepresented groups.

Stakeholder views: Royal College of Pathologists endorses hybrid models; patient groups like Roy Castle Lung Cancer Foundation welcome consistency. Future trials may assess clinical utility in multi-centre NHS cohorts.

Looking Ahead: Transforming UK Lung Cancer Care Through AI

As digital pathology matures, expect AI to evolve: multi-biomarker panels, real-time intraoperative guidance. For UK higher education, this heralds interdisciplinary hubs blending pathology, data science, and oncology.

Patients gain personalised care; researchers, new datasets. Visit ICR's announcement or the study at Scientific Reports for deeper insights.

This tool not only sharpens diagnostics but inspires the next generation of UK academics to harness AI against cancer.

Photo by Thorium on Unsplash