Google says UK breast cancer screening AI found 25% of missed interval cancers

On March 10, 2026, Google published new breast cancer screening results from work conducted with Imperial College London and the UK’s National Health Service. The studies, published in Nature Cancer, focus on one of the hardest operational questions in medical AI: whether an AI system can improve detection at scale without weakening the clinical safeguards that human specialists rely on. Google said its experimental research system identified 25% of the interval cancers that had previously been missed in screening, a result that directly targets cases that often surface only after symptoms appear and treatment becomes more difficult.

The first study compared AI-assisted mammography interpretation with expert radiologist performance using mammograms from 125,000 women. According to Google, the system not only recovered 25% of total missed interval cancers, but also identified more invasive cancers and more cancers overall than expert radiologists, while producing fewer false positives for women receiving a first-time scan. Those are meaningful operational claims because breast cancer screening programs are judged not only by detection yield, but also by the downstream burden of unnecessary recalls and follow-up procedures.

The second study asked a more practical deployment question: whether AI can relieve workforce pressure inside the NHS double-reading workflow. In the UK system, two specialists must agree on each mammogram, and an arbitration panel resolves disagreements. Google said the study covered scans from over 50,000 women and found that AI could reduce screening workload by an estimated 40% when used as the second reader. That finding matters because each specialist reviews roughly 5,000 scans annually, yet dedicated review time is limited and radiologist shortages remain a persistent bottleneck. If the estimate holds up in live deployment, AI could create capacity without lowering the formal review standard.

Google’s post is careful not to oversell the result. The company said specialists in arbitration sometimes overruled AI-detected cancers that otherwise would have remained undetected, highlighting that trust, workflow design, and human-AI interaction are still open problems. It also described an observational feasibility study across 12 NHS screening sites in London that processed over 9,000 cases in real time without affecting patient care, and concluded that clinical AI is not a plug-and-play product. Calibration to hospital workflows, equipment, and patient populations remains essential. Even so, the combination of detection gains, workload reduction, and real-world workflow data makes this one of the more concrete medical AI updates of the month.