AI second look helps NHS Grampian find 11 extra breast cancers

An AI tool helped screening staff in Aberdeen find 11 breast cancers that routine screening had not picked up. That is the main real-world result from a study at NHS Grampian, published in Nature Cancer on 10 March 2026.

In Scotland, two trained readers look at every screening mammogram. That did not change. Between 27 February and 5 October 2023, an AI tool called Mia, made by Kheiron Medical Technologies, also read the images. The study is called GEMINI, short for Grampian's Evaluation of Mia in an Innovative National breast screening Initiative. It has nothing to do with Google's Gemini AI.

When the AI said a woman should be called back but the human readers had not, senior readers took another look. People made every decision. They only saw the AI's opinion after the normal double reading was finished, so it could not sway the first reads.

The study covered 10,889 women. Routine screening found 106 cancers among them. The extra look prompted by the AI found 11 more. Seven were invasive cancers. Four were DCIS, early changes inside the milk ducts.

To get there, the AI flagged 1,345 women the humans had not recalled. After a second look, 55 of them were called back. Most of those extra checks were quick: 63% took 30 seconds or less.

This safety net was not free. Used as an extra check on top of normal screening, the AI raised the reading workload by 6%. It also raised the number of women recalled by 11.3%. More recalls mean more women asked back for tests, and more worry while they wait.

The researchers also used the AI's results to model 17 different ways it could be used, including set-ups that might cut work for staff. Those are computer simulations of what might happen. They were not tested on women in live screening, so they are not real-world results.

There was also a very ordinary problem. Images from 4,992 women, nearly three in ten, never reached the AI at all. An old IT coding error made sessions time out after 30 minutes of inactivity. The researchers checked and found that the human results did not differ significantly between women affected and those who were not. Still, it shows that plugging AI into an old hospital system is often the hard part.

Some limits are worth knowing. This was not a randomised trial. It was a service evaluation at one site, using one AI tool. Three of the nine authors worked for Kheiron and held stock options. The researchers say the analysis was done in a secure workspace kept apart from the company, which was given separate access to check the results. The study was funded by an NHS AI award.

Why is this a good-news story? The AI did not replace anyone. It acted as an extra pair of eyes, and trained people decided what happened next. That set-up found cancers that routine screening had not picked up, at the price of more work and more recalls that the study measured openly.