PD-L1 scoring in lung cancer: TPS-Guard sends pathologists the slides on which its 15 models disagree, a sound approach tested on one centre and one antibody

A team from the University of Warwick and Coventry's university hospitals presents a system that automatically computes the PD-L1 score (TPS) on non-small cell lung cancer slides, then sends to a pathologist, with a stated reason, the estimates on which 15 models disagree most. On 114 slides from a single centre that were not seen in training, the model's error (6.96 points) stays within the range of three pathologists (4.24 to 8.29), and its uncertainty flags the very slides on which pathologists disagree among themselves. The triage idea is sound, but the error triples on external data stained with another antibody, the alerts were tuned on the cohort used to evaluate them, and no study yet shows what they change in real reporting.

The context

In advanced non-small cell lung cancer (NSCLC), the choice of immunotherapy partly rests on one number: the tumor proportion score (TPS), the percentage of tumour cells whose membrane is stained, even weakly, by anti-PD-L1 immunohistochemistry (a stain that reveals the PD-L1 protein on the slide). Two cutoffs matter: 1% and 50%. Above 50%, pembrolizumab can be offered alone as first-line treatment.

The score is estimated by eye, and pathologists do not always agree: tumour cells must be told apart from macrophages, which can also stain, partial staining must be recognised, and the estimate extrapolated across heterogeneous tissue, sometimes a small biopsy. Several studies have shown that differences between readers can exceed differences between antibodies. Existing automated tools (Hondelink et al. 2022, van Eekelen et al. 2024) report average agreement across a cohort, without saying which individual estimate deserves a second look. That is the question of this preprint: can an algorithm flag its own doubts, and do those doubts coincide with the cases humans find hard?

The method

The system follows the steps of a human read. First, outline the tumour: digitised slides are cut into 224 × 224-pixel squares (patches), described by Virchow2, a pathology foundation model (a large network pre-trained on millions of tissue images and reusable for many tasks), then classified as tumour, non-tumour or macrophage-rich, the latter being excluded. Next, count: the D2Count-Net network has two branches, one detecting each cell (effective when cells are spread out), the other estimating a density map (more reliable when they are crowded), weighted according to tissue cellularity. TPS is the number of positive tumour cells divided by all tumour cells.

Finally, TPS-Guard: fifteen independently trained models each produce a TPS, and their standard deviation serves as the uncertainty measure. The 15% most uncertain slides are sent to the pathologist; the others are "auto-reported", with the pathologist remaining responsible for every report. Three alerts give a reason: possible over-scoring around 1%, possible under-scoring around 50%, or too few tumour cells analysed.

The data come from a single hospital, University Hospitals Coventry and Warwickshire: 154 slides from 152 patients (mean age 70, 65% women), stained with the SP263 antibody and digitised on a single Philips scanner. 45% have a TPS below 1% and 14% a TPS of at least 50%. Cell annotations cover 42 slides and 50,615 nuclei, only 6.2% of them positive tumour cells. Three pathologists (consultant and trainee level) scored each slide blinded; the reference is the mean of the three. The main analysis covers the 114 slides never used in training, and both the model and each pathologist are also compared with the mean of the other two readers. External validation uses the public IGNITE dataset (Nijmegen, the Netherlands, and Negrar, Italy): 98 regions from 21 patients, stained with E1L3N or SP263, on three other scanners.

The results

On the 114 slides, the mean absolute error (MAE, the average gap in percentage points from the reference) is 6.70 points, versus 6.41 points between two pathologists. Against the mean of the other two readers, the pathologists score 4.24, 5.49 and 8.29 points and the model 6.96: 0.95 points more than a single pathologist (95% confidence interval −0.53 to 2.48). The model underestimates high scores (mean bias −5.0 points).

The central result concerns uncertainty. It correlates with model error (Spearman correlation ρ = 0.77) and with disagreement between pathologists (ρ = 0.71). Mean error rises from 0.6 points in the least uncertain quarter of slides to 16.4 points in the most uncertain quarter. After referring 17 slides (14.9%), the error on the remaining 97 falls to 4.96 points, matching the gap between pathologists on those same slides (4.93).

At the clinical cutoffs, the picture is more mixed. At 50%, the model misclassifies 5.6% of slides, versus 6.1% for a single pathologist; at 1%, 19.9% versus 17.5%. Sensitivity at the 50% cutoff (the share of slides truly ≥50% recognised as such) is 10 out of 16, and triage does not improve it: it refers four of the six false negatives, but also eight of the ten true positives. In practice, by our calculation on these small numbers: out of 1,000 slides with the same profile, about 140 would be above 50% and the model alone would place about 50 of them below, as many patients for whom single-agent immunotherapy might not be discussed without a second read.

Externally, without retraining, the MAE climbs to 21.12 points: 11.06 on regions stained with SP263, the development antibody, but 29.66 on E1L3N, never seen in training. Fine-tuning (partial retraining on local data) only brings it down to 19.05. Uncertainty ranking remains useful: referring 10% to 50% of regions lowers the error of the remaining cases from 20.5 to 12.4 points.

What is good

First, it asks the right question. Rather than claiming to replace the pathologist, the paper asks when the automated number can be trusted, and shows it with a simple measure: disagreement among 15 models predicts both machine error and human divergence. Each alert comes with a readable reason, which makes it a review tool rather than a black box.

Second, the comparison with readers is fair. The model is judged by the same protocol as each pathologist, against the mean of the other two, with confidence intervals, and results with the original scores (three scores were corrected afterwards) are published in the supplement.

Third, the bad news is in the main text: the collapse on E1L3N, the weak sensitivity at 50%, the fusion of the two branches being beaten by detection alone after adaptation, the annotation imbalance. The external dataset is public, which allows independent comparison.

What is less good

First, population bias in the broad sense. One hospital, one antibody, one scanner, 114 evaluation slides: that is narrow for a tool meant for laboratories using other antibodies (22C3, E1L3N) and other scanners. External validation confirms it: the error triples, and because antibody, centre and scanner vary together, we cannot tell which one is responsible. The development data are not public and the code will only be released after acceptance.

Second, information leakage in tuning the alerts. The rules and the 15% threshold were defined on the cohort where their effectiveness is measured; externally, some thresholds were recomputed from the model's outputs on the test set itself. The reported enrichments (33% false positives among slides flagged at the 1% cutoff, versus 11.4% on average; 4 of 8 slides at the 50% cutoff) are therefore optimistic, and the authors call them exploratory.

Third, a misleading metric. Nearly half the slides are below 1%, where absolute error is mechanically small: the average MAE flatters the model where the weightiest decision is made, above 50%, where sensitivity is only 0.63 on 16 slides. No reader study with and without the tool was conducted. On interests, co-author David Snead is a shareholder and director of Histofy Ltd, a digital pathology company; the declared funding is a Warwick PhD studentship. Virchow2 is distributed under a non-commercial licence (CC BY-NC-ND 4.0), which would complicate commercial use as is.

What it changes

For research, the work usefully shifts the question from "what is the average error?" to "which estimate should be checked again?". The link between model uncertainty and human disagreement applies beyond PD-L1, to other semi-quantitative pathology scores such as Ki-67 or HER2. The announced next step is the right one: rules fixed in advance, new slides, several centres, pathologists with and without alerts.

For clinicians, nothing changes today: this is a research prototype, with no regulatory status, evaluated retrospectively. Above all, the study is a reminder that human TPS is itself uncertain near the cutoffs, and that a tool that underestimates high scores would be dangerous without review.

For patients, the long-term promise is a more reproducible score and human attention focused on genuinely ambiguous cases. But in a laboratory that does not use Coventry's antibody and scanner, this system would often be wrong as it stands and would first need local validation.

Further reading

The preprint: Automated PD-L1 Tumor Proportion Scoring with Uncertainty-Guided Triage in Non-Small Cell Lung Cancer, Xu, Azam, Kaur et al., University of Warwick (Tissue Image Analytics Centre), University Hospitals Coventry and Warwickshire, University Hospitals Birmingham, Royal Derby Hospital and Donghua University, arXiv (cs.CV), submitted 12 November 2026, DOI 10.48550/arXiv.2611.16907. Not peer reviewed.

External data: IGNITE data toolkit on Zenodo. Base model: Virchow2 on Hugging Face.

On Tatakoto: slide quality, a prerequisite for any AI in pathology and LungAI and lung nodule reclassification.

Editorial transparency note: this article was produced by Tatakoto's automated editorial pipeline from version 1 of the preprint (arXiv PDF, main text and supplement). The "out of 1,000 slides" conversion is our calculation from the published proportions and rests on small numbers. The English, Spanish and Chinese versions are AI-generated. Spotted an error? Contact us.