A pathology AI ensemble recognised Barrett’s cancer—but has not yet improved care
A peer-reviewed benchmark combined five pathology foundation models across magnifications. It performed strongly on 3,528 archived slides, but remains a retrospective classifier rather than a tested clinical diagnostic.
The short version
- What happened
- The benchmark used 3,528 digitised tissue slides from 790 patients, covering tissue from normal appearance through Barrett's changes to oesophageal adenocarcinoma.
- Who it affects
- For people undergoing Barrett's surveillance, a reliable second-read tool could eventually help pathology teams prioritise suspicious samples. This paper does not show that it reduces missed cancer, unnecessary treatment or waiting time in routine care.
- What remains unproven
- The study was retrospective and slide based; it did not prospectively evaluate pathologists working with the tool.
Keep in mind
The study was retrospective and slide based; it did not prospectively evaluate pathologists working with the tool.
Reported accuracy and AUROC are not patient outcomes, and overall accuracy can obscure weaker performance in rarer morphology classes.
The material used a particular fixation and digitisation workflow, so external sites need validation across their own laboratories and scanners.
No effect on diagnostic decisions, treatment, workload, safety or survival was measured.
Source analysis · 2 linked sources. Primary analysed source date: 21 October 2026. Earliest linked supporting source: 11 November 2025.
Read the full report · 3 min The evidence, methods, implications and limitations
Combining views helped more than trusting one model
The team tested pathology foundation models on haematoxylin-and-eosin slides viewed at magnifications from 5× to 40×. It then combined the five single-scale models whose predictions agreed most closely with experts. The aim was to capture both broad tissue structure and cellular detail rather than rely on one representation.
For cancer detection, the ensemble reported an area under the receiver-operating curve of 0.909, an F1 score of 0.836 and accuracy of 0.959. The morphology task was harder: its corresponding figures were 0.907, 0.696 and 0.795, with kappa agreement of 0.651.[1][2]
Accuracy did not mean every category was equally reliable
The denominator matters. The dataset contained 3,528 whole-slide images from 790 patients and represented stages from normal tissue through metaplasia and dysplasia to cancer. Overall accuracy can look high when some categories are more common or easier, which is why the lower macro F1 score for morphological subtyping deserves equal attention.
Agreement with pathologists is useful for prioritising suspicious slides, but it is not the same as proving a safer diagnosis. The study did not randomise clinical readers, compare final treatment decisions or test whether a missed or overcalled category changed what happened to a patient.[2]
The distinction that matters
From one tissue slide to an ensemble decision
The system aggregates information across image scales and model backbones before producing a slide-level classification.
Digitise
Turn each tissue slide into image tiles
3,528 slides from 790 patients
Compare
Run models from 5× to 40×
Different tissue and cellular views
Ensemble
Combine five expert-consistent models
One slide-level prediction
A benchmark, not a deployment verdict
The experiment used archived, specially fixed and digitised slides. Performance transferred across fixation conditions without task-specific fine-tuning in the reported tests, but a pathology service would still need external validation on its scanners, staining, specimen mix and workflow before relying on the model.
The authors declared no competing interests in the public preprint. The next persuasive evidence would be a prospective, multi-centre reader study with pre-specified error categories, calibration and patient-relevant outcomes. Until then, the ensemble shows that model diversity can improve a benchmark—not that it can replace expert pathology review.[1][2]