Prof Gerald Lip, Clinical Director for breast screening in the North-East of Scotland and the Principal Investigator of the GEMINI evaluation, answers some key questions about the evolving role of AI in breast screening
Artificial intelligence (AI) is increasingly part of the conversation about the future of breast screening. The GEMINI evaluation, carried out in Scotland, offers valuable insights into how AI can support, rather than replace, clinical expertise.
What was the GEMINI evaluation?
GEMINI was an evaluation designed to understand the impact of using AI in breast screening as a safety net alongside standard practice. It took place in NHS Grampian, Scotland, between February and October 2023.
Around 11,000 women due to attend routine breast screening were invited to take part. All participants received the same standard of care: Their mammograms were read independently by two radiologists, with a third reader involved if there was disagreement.
The difference was that, where a mammogram was given a normal result by human readers, AI was switched on in the background to check this reading. If the AI highlighted an area of concern, radiologists reviewed the images again.
What did the evaluation find?
Using AI in this way led to the detection of 10 per cent more breast cancers. These additional cancers were generally small and early-stage – the type that might not have become visible until the woman’s next screening mammogram.
Importantly, these findings were subtle. None of the cancers were obvious on first reading and all could reasonably have been not seen by human readers alone. The AI acted as an additional safety net rather than a replacement for clinical judgement.
Similar results have been seen elsewhere. For example, the Swedish MASAI trial of 100,000 women found that when AI was used to support screening, 29 per cent more cancers were detected. In Denmark, AI was introduced to replace one human reader as a solution to recruitment challenges which were causing delays. They screened as usual for six months with two human readers then introduced AI and screened for another six months with one human reader and one AI reader. In the six months with one human and one AI reader they picked up more cancers, had fewer false positives, and reduced the reading workload.
How did women involved in GEMINI feel about AI being used?
Women were largely supportive. Of approximately 15,000 invitation letters sent, only 124 women chose to opt out. Those who did shared a range of concerns, including a general dislike of AI, worries about data use, or fears that radiologists might lose their jobs.
Clear communication helped. Providing FAQs and transparent information reassured many women that AI was being used in the evaluation safely and responsibly and always with humans in the loop.
Did AI increase recall rates?
Yes, AI flagged around 1,000 additional cases for review. However, expert human readers carefully assessed these and ultimately recalled only 55 women. From those 55 recalls, 11 additional cancers were diagnosed.
This is why humans will always be needed. AI can generate false positives, which leads to an increased workload for radiologists. It takes skilled professionals to interpret, filter, and act on its results. The most benefit comes from combining AI with human expertise.
So AI will not replace human readers?
No. AI does not replace radiologists – and it cannot. AI analyses images only. It does not consider prior mammograms, clinical and screening history, or individual risk factors.
Humans provide context, judgement, and experience. We can recognise when an anomaly flagged by AI has been stable for years, previously biopsied, or can be explained by clinical information that AI cannot access.
The strongest results come from human–AI collaboration. AI can reduce workload, support consistency, and help prevent the burnout that can be caused by reading a high volume of mammograms and carrying the responsibility for women’s health. It gives radiologists more time and space to apply their expertise.
What are the benefits for women?
One major benefit is faster turnaround times. AI provides an instant second read, rather than waiting for another human reader.
Faster results mean quicker follow-up appointments, reduced waiting times, and less anxiety for women – all contributing to a better screening experience.
There is also growing research interest in AI’s potential to support risk prediction and personalised screening, with early evidence suggesting AI may identify
women who could benefit from earlier recall between screening rounds.
What are the common misconceptions about medical AI?
One common misconception is that current medical AI is ‘learning’ continuously like some AI systems, when it is actually a fixed software version. Medical AI used in screening is CE-marked software, which means it meets regulatory, safety, and performance requirements. This type of AI delivers fixed, repeatable outputs, while performance can be monitored and calibrated. It does not independently evolve as it is used.
AI is a tool – not a decision-maker. It supports clinicians and makes us better, acting as a reliable ‘second pair of eyes’ or ‘friend over your shoulder’, helping us feel more confident in our decisions.
What is next for AI in breast screening?
Sweden used to double read, but it is about to switch to having one human plus AI. In the Swedish MASAI Trial, AI-supported breast screening detected 29 per cent more breast cancers, while potentially halving
radiologist workload.
Seven EU countries are already using AI in radiology in practice. Norway and the UK are trialling AI. The UK trial is over three years and its findings will inform future decisions. Countries where there is single reading are already using AI.
In the US some people are paying under their health insurance policies to have AI added to their reading. They are paying for a single reader plus AI.
Radiology is leading the way, but lessons learned here will support AI adoption across other clinical specialties, including pathology, scheduling, reporting, and patient communication.
There is also potential for research because it offers a good amount of data to harvest, enabling us to look at prediction and personalisation.
How can AI be introduced safely into screening programmes?
One effective and safe approach is ‘ghost or silent running’, which means running AI in parallel with routine screening for one to three months without it influencing clinical decisions. This allows services to assess its performance, understand the rate and nature of false positives, and build confidence before making AI live.
What matters most is calibration; fine-tuning AI to reflect the characteristics of the population being screened.
What does calibrating AI to different populations mean?
Screening populations differ between regions, meaning that AI needs to be adjusted to account for variations in cancer incidence, age profiles, ethnicity, and historical risk factors.
For example, populations with different ethnic make-ups, or known historical exposures, may have different baseline cancer risks. Calibration allows AI performance to be optimised for those local characteristics, improving accuracy and relevance.
What about bias and consistency?
All AI models have inherent biases based on how they were developed, but, unlike humans, AI performance is consistent. Human readers vary; some are more likely to recall a woman than others. AI can be calibrated to sit somewhere in the middle, providing stability.
Over time, clinical teams learn how to work effectively with AI. Safeguards are needed to avoid automation bias, where humans rely too heavily on AI, or even learn from it. Maintaining independent reads remains essential.
Final thoughts
Clinically, the evidence is clear: AI works when used responsibly. However, introducing AI into breast screening is complex. It requires robust infrastructure, governance, and careful implementation.
AI is already helping detect cancers today that might otherwise have gone undetected for another two or three years. Used well, it has the potential to strengthen screening programmes, support staff, and most importantly, improve outcomes for women.
Further information: www2.healthservice.hse.ie/organisation/nss/news/advancing-screening-in-ireland-implementing-an-ai-programme-in-nss/ or
www2.healthservice.hse.ie/organisation/nss/news/artificial-intelligence-and-population-screening/
GEMINI: www.nhsgrampian.org/about-us/innovation-hub/our-projects/previous-projects/gemini/
Leave a Reply
You must be logged in to post a comment.