AI in NHS breast screening

Summary

• Breast screening sends women for a mammogram, a picture of the breast, to find early cancer. Two radiologists read each one, a practice called double reading, and the government says two specialists are needed per mammogram screening.
• The Royal College of Radiologists says the UK has 32% fewer radiologists than it needs, and projects a 40% gap by 2030 if nothing changes.
• Supporters argue that one reader with AI could match two and free up staff time. Critics answer that a second human reader may catch what the AI misses, and that an AI which flags more may send more women for tests they do not need.
• The UK evidence is the GEMINI study of Kheiron Medical's tool Mia, on 10,889 women in one region. Where Mia recommended a recall that double reading had not, extra human review found 11 more cancers. The paper reports a 10.4% relative rise, or one extra cancer per 1,000 women.
• Sweden's MASAI trial gave about 106,000 women AI supported reading. Its 2026 Lancet report tested whether AI was no worse for interval cancers, the ones found between screens. It found 82 interval cancers with AI and 93 without, and AI cut screen reading workload by 44%.
• EDITH compares double reading with two ways of adding AI, at about 30 centres, targeting 660,000 women. Recruitment began on 1 May 2026 and runs to 30 November 2028, so results are years away.

Dig deeper

• The 2023 safety analysis in The Lancet Oncology randomised 80,033 women: 6.1 cancers per 1,000 with AI against 5.1 without, not statistically significant (p=0.052).
• Women are randomly allocated and will not know which group they are in. Mammograms and patient information go to CIMAR (UK) Ltd for AI reading, which holds them for up to 30 days.
• Using identifiable records without consent needs special approval, which EDITH has after advice from the Confidentiality Advisory Group.
• The Medical Protection Society warns that clinicians may absorb all legal responsibility, because the Consumer Protection Act 1987 probably does not cover AI software.
• The question of who may hold patient data for a private company also arises in the Synnovis NHS cyber attack.

Ethical issues

Interview questions

• What do you know about the EDITH trial?
• What did the GEMINI study find when Mia was tested in one UK region?
• You are the second reader on a mammogram. The AI has flagged nothing, and you can see a faint shadow. What do you do?
• Why might finding more cancers still leave a woman worse off?
• Should a woman be told which reading group she is in before she agrees to a scan?
• If an AI tool misses a cancer that a radiologist would have found, who should be held responsible, and why?
• Should public money fund a trial of a tool made by a private company, and what is the case against it?
• Why would a trial of this kind need several years before it can show whether fewer women die?