Google AI catches 25% of breast cancers missed by NHS screening

TL;DR:

• An AI system developed by Google and Imperial College London identified 25% of interval cancers previously missed by NHS radiologists, across a study of 125,000 mammograms published in Nature Cancer.
• A second study of 50,000 women showed AI could reduce screening workloads by an estimated 40% when used as a second reader, directly addressing the UK’s radiologist shortage.
• A feasibility study across 12 NHS sites in London found that AI is not a plug-and-play solution and requires continuous calibration to each hospital’s equipment and patient population.

Two studies published in Nature Cancer present the strongest evidence yet that AI can improve breast cancer detection within the NHS. The research, a collaboration between Google, Imperial College London, and NHS screening sites, tested an AI system against the UK’s existing double-reading process where two specialists must agree on every mammogram.

The headline finding: across 125,000 mammograms, the AI identified 25% of interval cancers that had been missed. These are the cases that slip through routine screening and only surface when symptoms appear, typically at a more advanced stage.

Practical workload relief

A companion study examined whether AI could function as a second reader, effectively replacing one of the two human specialists required for every scan. Across 50,000 cases, the researchers estimated a 40% reduction in screening workload. Each NHS specialist currently reviews around 5,000 scans annually with only four hours of dedicated screening time per week, so a meaningful reduction would free capacity for complex cases and help address the nationwide screening backlog.

Trust remains the bottleneck

The research also exposed a tension that technology alone cannot solve. During simulated reviews, arbitration panel specialists occasionally overruled AI-detected cancers that would have gone undetected without the system. In other words, radiologists sometimes rejected correct AI findings, a trust gap that the researchers say requires further work on human-AI interaction.

An observational feasibility study across 12 London NHS screening sites reinforced the implementation challenge. Processing over 9,000 cases in real time without affecting patient care, the study found that each hospital required individual calibration, continuous adjustment, and adaptation to local equipment and patient populations.

These findings build on earlier Google-NHS research showing AI-based screening could shorten diagnostic waiting periods. The progression from lab accuracy to real-world deployment studies marks an important step, but the trust and calibration challenges suggest widespread NHS adoption remains some distance away.