Google, Imperial College and the NHS
AI second reader finds 25% of interval breast cancers missed by screening (two Nature Cancer papers)
Result confirmed
The takeaway
On March 10, 2026 Google Research, Imperial College London and the UK’s NHS published two studies in Nature Cancer on an experimental AI system for breast-cancer screening.
Status
- Claim
Result confirmed
- Our reporting
- Medium confidence
- Verification
- Unverified claim
- Importance
- 3 of 5
- Last verified
- 13 October 2026
Your AI and this story
- GPT-6 AstraIn its training data
- Claude Opus 5.5In its training data
- Gemini 3.8 FlashIn its training data
- Grok 4.7In its training data
It happened before the cutoffs of all four assistants, so it can be in their training data.
Key facts
- Two papers in Nature Cancer: s43018-026-01127-0 (accuracy) and s43018-026-01128-z (workload and arbitration); the papers themselves were not readable by our fetcher (login redirect), so numbers come from Google’s posts and trade press
- Accuracy study: mammograms of 125,000 women; the AI detected 25% of interval cancers (cancers diagnosed between screening rounds after a normal result) that had been missed; more invasive cancers and more cancers overall than expert radiologists; fewer false positives at first-time screens (Google, Mar 10)
- ICT&health (Mar 11): about 125,000 women aged 50–70 from five NHS screening programmes, screened 2015–2016; 115,973 scans in the final analysis; detection rose from 7.54 to 9.33 cancers per 1,000 women with AI as second reader; for first-time screeners recalls fell 39.3% and detection rose 8.8%
- Workload: Google says an AI second reader was estimated to reduce screening workload by 40% (study of over 50,000 women); ICT&health reports a 32.1% fall in scan-reading workload in the retrospective study. The two figures are not reconciled here
- Real-time feasibility study: over 9,000 cases (ICT&health: 9,266) across 12 NHS screening sites in London; AI output did not influence patient care; AI took 17.7 minutes per assessment on average vs 2.08 days for the first human reader (ICT&health)
- Arbitration: in a simulated review, specialist arbitration panels sometimes overruled cancers that the AI had correctly flagged (Google)
- Context: NHS screening uses double reading; each specialist reads about 5,000 scans a year (Google). The UK has a 29% shortfall of clinical radiologists (ICT&health)
- Oct 12, 2026: Google’s roundup ‘Applying technology to understand and treat disease’ (Yossi Matias, Pushmeet Kohli) repeats: 125,000 women, 25% of missed interval cancers, workload reduction ‘up to 40%’
What happened
The NHS breast screening programme has every mammogram read by two specialists. Google’s experimental system was tested as a replacement for the second reader. On historical scans of about 125,000 women it found a quarter of the interval cancers, the ones that appeared between screening rounds after a normal result. A companion paper modelled the workflow and estimated a large cut in reading work. A small real-time run at 12 London sites showed the system can return results in minutes, without affecting care.
Google’s Oct 12, 2026 roundup of its health research (90 peer-reviewed papers in five years, by its count) lists these studies next to AlphaFold drug repurposing for Chagas disease, AlphaGenome Atlas, AlphaProtein Novo, DeepSomatic and a flu forecast ranked first by the CDC for the 2025–26 season. Google.org says it gave over $60 million in 2026 to groups applying such models to disease.
Why it matters
Interval cancers are the main failure of screening, and the UK lacks radiologists. The result is mostly retrospective: whether an AI reader improves outcomes in live screening is being tested in prospective trials (ICT&health mentions the EDITH trial across 30 sites). The arbitration finding is a warning that humans may overrule correct AI calls.
What is disputed or not yet verified
The Nature Cancer papers were not read directly; titles, author lists and exact publication dates of each paper are not confirmed. The 40% and 32.1% workload figures come from different sources.
Sources
5 sources from 3 sites. Numbers match the chips in the text.
5 sources: 4 primary, 1 press
Primary
- Google: How AI can improve breast cancer detection in the UK (Mar 10, 2026)blog.google, official
- Nature Cancer paper 1 (s43018-026-01127-0)nature.com, paper
- Nature Cancer paper 2 (s43018-026-01128-z)nature.com, paper
- Google: Applying technology to understand and treat disease (Oct 12, 2026)blog.google, official
Press
Changes
- Filed as a review (written from Google’s two posts and ICT&health)