News · 30 July 2026

Digital slides at national scale: what 15 connected NHS sites mean for pathology AI

In July 2026 the National Hospital for Neurology and Neurosurgery, part of University College London Hospitals, joined the UK's national digital pathology network, bringing the number of connected NHS sites to 15. The network is run by the National Pathology Imaging Co-operative, based at Leeds Teaching Hospitals NHS Trust. It lets pathologists view and analyse high-resolution digital slides remotely, request second opinions without shipping glass, and, over time, apply AI-based tools. More hospitals are expected to join later in 2026.

Why scale matters

A single hospital's slide archive reflects its own staining protocols, scanners and patient population. Models trained there often perform worse elsewhere. A shared platform across many sites makes it possible to test whether a model holds up across institutions, which is exactly the evidence regulators and clinicians want.

The technical challenges that remain

Domain shift. Differences in stain, scanner and tissue preparation can change an image's appearance enough to degrade a model. Stain normalisation, augmentation and multi-site validation help, but they must be measured, not assumed.

Gigapixel images. A single slide contains billions of pixels. Models must process them in tiles and aggregate results, which creates real choices about where an explanation should point.

Weak labels. Many slides come with a diagnosis for the whole case, not an annotation of each region. Learning from such labels is hard, and an explanation helps check that a model has learned morphology and not an artefact, such as a pen mark or a scanner signature.

Workflow fit. A tool that adds minutes to each case will not be used, however accurate it is.

What explainability adds in pathology

For a pathologist the useful question is not "what is the probability of cancer?" but "show me what you are looking at." Heatmaps over regions that drove a prediction, retrieval of similar regions with known diagnoses and indications of uncertainty give an expert something to confirm or contradict.

These aids also support quality assurance. If a model's attention concentrates on unlikely areas, that is a warning that something is wrong with the data or the model.

A practical checklist for developers

  1. Validate on data from sites and scanners not used in training.
  2. Report performance by subgroup and by scanner.
  3. Test explanations with pathologists, comparing decisions with and without them.
  4. Monitor after deployment, and keep versioned records.

Digital infrastructure is the precondition for pathology AI. Rigorous validation and explanation are what turn it into something clinicians can rely on.

Written in July 2026; details reflect the situation at that time.

All news