Google, Imperial, and the NHS say AI catches 25% of interval breast cancers missed in screening

Google said on March 10, 2026 that joint research with Imperial College London and the UK’s National Health Service could improve mammography screening in the NHS. The announcement is based on two studies published in Nature Cancer and a feasibility exercise across London screening sites, giving the release both a research and workflow operations angle.

What the studies found

In the first study, researchers used an experimental AI system to review mammograms from 125,000 women. Google says the system detected 25% of interval cancers that had been missed in routine screening. The company also says the AI identified more invasive cancers and more cancers overall than expert radiologists, while producing fewer false positives for women undergoing first-time screening.

The second study evaluated more than 50,000 women and looked at AI as the second reader in the NHS double-reading workflow. In that setting, Google estimates screening workload could fall by 40% while preserving the existing clinical structure. That matters because the NHS model depends on agreement between two specialists and operates under sustained staffing pressure.

More than an accuracy story

The most important part of the announcement may be what happened when clinicians and AI disagreed. Google says arbitration panel specialists sometimes overruled cancers flagged by AI during simulated review, even in cases that might otherwise have remained undetected. That suggests deployment risk is not only about model accuracy, but also about trust, escalation design, and how AI findings are presented to human readers.

The observational feasibility work covered 12 NHS screening sites in London and processed more than 9,000 cases in real time without affecting patient care. Google says the exercise showed that AI is not plug-and-play in healthcare settings and requires local calibration to workflow patterns, imaging equipment, and patient populations.

Why it matters

Breast cancer screening in the UK relies on a demanding double-reading process and a limited specialist workforce. A system that can surface additional cancers and reduce reading burden could help address backlogs, but the Nature Cancer results do not automatically translate into immediate nationwide deployment. The more immediate takeaway is that health AI is moving beyond retrospective benchmark claims toward workflow studies that test how clinicians and models behave together under real service conditions.

Source: Google