(Published in the UCL IMS Magazine – Edition 2 – March 2026)
From adversary to ally…
AI is everywhere – you can’t escape it. These large language models have become the byword for theft, job-stealing, environmental pollution and of course, cheating on homework, with the overwhelming response largely negative. Its propensity for medical misinformation has worried the medical population, especially given the plethora of it out there in a post-COVID population less trusting of conventional medicine than before.
Practically no one, therefore, would have said that these models could actually help the medical field, particularly in that most challenging of areas – oncology. And yet, an abundance of studies have recently shown that AI has promise in improving breast cancer detection by working alongside radiologists, mainly in assessing patient cancer risk and encouraging Radiologists to reassess high-risk cases to triple-check.
This has been of much interest, especially in the detection of interval cancers.
Interval Cancers:
Interval cancers are cancers diagnosed after a screening test gives a negative result but before the next screening. The cancer either appears for the first time or becomes clinically detectable in the interval.
They are associated with biological features that lead to faster growth, greater aggressiveness and a worse prognosis.
Why Mammograms Miss Tumours:
This can be due to several factors:
- False negative screening result (aka ‘missed’ cancer) where the cancer was present but:
- Too small to detect
- Obscured
- Misinterpreted by the radiologist (this is key – could AI help?)
A fast-growing tumour which grows rapidly between screenings. It was not present at the last screening and only becomes detectable at the next.
A new cancer develops between screenings. It arises ‘de novo’ – which is common when the gaps between screenings are long.
In the UK, women are invited for breast cancer screenings every 3 years.
‘Double readings’ are standard in the UK and some Scandinavian countries: 2 independent radiologists review the same mammogram. This improves detection rates and reduces the chance of a cancer being missed.
However, while greatly improving accuracy and detection rates, it is time-consuming and affected by radiologist shortages. Therefore, lower-risk mammograms usually have only a single reading, which may become an issue if cancers are missed.
The Challenge of Breast Density:
Mammography sensitivity is greatly affected by participant age.
Younger women tend to have denser breast tissue (more difficult to read), whereas older women generally have fatter breast tissue, which is easier to read. Breast density can therefore affect the chances of cancer being caught, with the denser the tissue, the easier it is for a tumour to be ‘masked’.
Hence, women with denser breast tissue are usually assigned the double reading method as their scans are considered higher risk or more likely to be misread. Hence, AI will need to replicate the skill and experience of the radiologists who know how to adapt to reading mammograms with different presentations.
Can AI Replace a Second Radiologist?
Many papers in the last couple of years have aimed to assess whether AI is accurate enough to replace one radiologist in the ‘double reading’ method.
Frazer et al. (2024) found that AI performs as well as a single radiologist (single-reading method) in breast cancer detection but is inferior to the standard ‘double reading’ method. However, it did concede that the combined approach of human and AI proved to be the most effective.
This was shown in 2 methods:
AI ‘cleared’ many cases which were clearly shown to be ‘normal’ or noncancerous, hence drastically reducing radiologist workload and stress. AI acting as a ‘second reader’ – both the radiologist and AI read the mammogram and come to their conclusions separately, to avoid the radiologist being biased by AI’s decision. Therefore, AI here would take the place of the second radiologist.
These strategies were shown to increase the ‘sensitivity’ (how many of the patients with cancer were detected) of the scan by 1.9-2.5% and ‘specificity’ (how many healthy people are confirmed to be healthy) by 0.6%. The latter is especially important as it helps to rule out false positives and so avoid unnecessary stress for healthy patients for whom cancer is suspected. On a more economic note, it also means fewer biopsies need to be paid for and less (unnecessary) work for laboratory workers.
Automation Bias: A Hidden Risk
However, aside from this, the study makes an interesting point about ‘automation bias’ – where humans overly trust machines and systems due to the perception that they are more trustworthy. How many of us trust AI or Google over our own better judgement?
In this paper, automation bias was shown by radiologists mistrusting and changing their conclusions after seeing AI’s opinion and in the case of double-reader systems (2 radiologists), it was harmful. It removed the 2 independent opinions which had made this system so accurate, as the radiologists were more likely to agree with AI and hence with each other. Therefore, the recommendation is for the AI and the radiologists to read independently and come to their own conclusions.
However, in single-reader systems (only one radiologist), AI was beneficial. Both humans and AI models make mistakes, but they make different mistakes. AI can point out any things the human may have missed, and the human considers both submissions before ultimately making the final decision.
New Evidence: AI Detects More Cancers Than Double Reading:
A more recent paper by Gommers et al. (2026) also found that AI could be beneficial and detected more cancers than the much-lauded double-reader method. This was the case across all age groups and breast density levels, demonstrating AI’s ability to detect tumours that are often ‘masked’ by dense tissue.
As with Frazer et al, it also did not increase false positives and had an interesting effect on interval cancer detection. As stated above, Interval cancers are likely to be fast-growing, more aggressive and more invasive and hence often pose greater dangers than screen-detected cancers to the patient. Therefore, a worthy screening system should aim to reduce the number of interval cancers, detecting most cases during screening.
In this study, the AI group seemed to detect more cancers at screening, which would otherwise have been missed and developed into aggressive interval cancers – exactly what we want! Therefore, this shows that AI can help us detect cancers that would otherwise ‘slip through the net’ and reappear in a more dangerous form later. Given the 3-year gap between breast cancer screenings in the UK, this can only be beneficial.
The Limits of Current Research:
While the results from these and many other papers are promising, we must remember that they are mostly ‘non-inferiority’ trials, i.e., they confirm that the use of AI was not worse than the standard method.
While this was confirmed, more ‘superiority’ trials (to show AI is better than the current method) will need to be commissioned. The Radiologist shortage may prove to be a key contributor alongside positive studies to this implementation, but only time will tell. Who knows? Maybe one day the thing helping with your homework will one day save your life.
References:
- Lapid, N. (2024). Medical misinformation more likely to fool AI if source appears legitimate, study shows. Reuters. [online] 9 Feb. Available at:
https://www.reuters.com/business/healthcare-pharmaceuticals/medical-misinformation-more-likely-fool-ai-if-source-appears-legitimate-study-2024-02-09/(reuters.com in Bing). - Gradon, K.T. (2024). Generative artificial intelligence and medical disinformation. BMJ, [online] 384, p5979. doi: https://doi.org/10.1136/bmj.q5979.
- Shaw, M.L. (2024). Erosion of Trust in Health Care: A Public Health Crisis. [online] AJMC. Available at:
https://www.ajmc.com/view/erosion-of-trust-in-health-care-a-public-health-crisis(ajmc.com in Bing). - Performance Health Partners. (2023). The Hidden Insights 5 Years Post-COVID: Decline in Trust in Physicians and Hospitals. [online] Performance Health Partners. Available at:
https://www.performancehealthus.com/blog/trust-in-healthcare(performancehealthus.com in Bing). https://www.aladlam.com/newsroom/jchciar-adlam-slam-blog/the-human-impact-of-covid(aladlam.com in Bing)https://news.northwestern.edu/2024/08/07/trust-in-physicians-hospitals-research/(news.northwestern.edu in Bing)- England, N. (2020). Understanding Google AI on Breast Cancer: An overview of research conducted using AI to detect cancer in breast scans. [online] Imperial College London. Available at:
https://www.imperial.ac.uk/news/2020/new-research-conducted-using-google-ai-detecting-cancer-in-breast-scans/(imperial.ac.uk in Bing). https://blog.google/innovation-and-technology/health/google-ai-breast-cancer-detection/(blog.google in Bing)https://medicalxpress.com/news/2026-03-breast-cancer-uk.html(medicalxpress.com in Bing)- Díaz, O., Rodríguez-Ruiz, A., and Sechopoulos, I. (2024). Artificial Intelligence for breast cancer detection: Technology, challenges, and prospects. European Journal of Radiology, 175, pp.11457–11457. doi:
https://doi.org/10.1016/j.ejrad.2024.11457(doi.org in Bing). - England, N. (2020). What happens explained in the NHS Breast Screening Programme. [online] GOV.UK. Available at:
https://www.gov.uk/government/publications/breast-screening-programme-an-overview/what-happens-explained-in-the-nhs-breast-screening-programme-for-afab-assigned-patients(gov.uk in Bing). - NHS Breast Screening Programme. (n.d.). Available at:
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/983058/informed_consent_explained_in_breast_screening.pdf(assets.publishing.service.gov.uk in Bing). https://www.cancer.gov/types/breast/intraductal-cancer/define/intraductal-breast-cancer(cancer.gov in Bing)- Gommers, J., Hemström, V., Josefsson, V., Sartor, H., Schmidt, D., Hjelmgren, A., Larsson, A.-M., Hofvind, S., Andersson, I., Rosso, A., Högberg, O., and Ling, S. (2024). Interval cancer, sensitivity, and specificity compared with AI-assisted mammography screening: findings from the AI-MASAI study: a multicentre, randomised, non-inferiority, single-blind, population-based, scenario-driven study. The Lancet, 407(10257), pp.563–581. doi:
https://doi.org/10.1016/S0140-6736(24)00264-X(doi.org in Bing). - NHS (2024). Breast screening: guidance for image reading teams. [online] GOV.UK. Available at:
https://www.gov.uk/government/publications/breast-screening-guidance-for-image-reading/breast-screening-guidance-for-image-reading-teams(gov.uk in Bing). - Houssami, N., Lee, C.I., Buist, D.S.M., and Melnikow, J. (2017). Comparative sensitivity of mammography and tomosynthesis of double reading in digital mammography screening: A systematic review and meta-analysis. Breast, [online] 36, pp.74–81. doi:
https://doi.org/10.1016/j.breast.2017.09.006(doi.org in Bing). https://medicalxpress.com/news/2020/09/trained-radiographers-perform-as-well-as-radiologists.html(medicalxpress.com in Bing)- Stabling, S. (2024). Breast cancer screening: NHS process. [online] doi: https://doi.org/10.46628/jr.2024.6.
- Helen, Parida, S., C.A., Kwok, C.F., Elliott, M.S., Glenn, Y., Wang, C., Osamah Al-Qarshi, R., S.K., Hill, E., Revalation, K., S.L., Ellis, L., and Mottick, N. (2024). NHS Breast Screening Programme: Benefits, limitations, and informed consent. [online] NHS Breast Screening Programme. Available at:
https://www.gov.uk/government/publications/breast-screening-helping-you-decide(gov.uk in Bing). - NHS (2024). NHS Breast Screening: Helping You Decide. [online] GOV.UK. Available at:
https://www.gov.uk/government/publications/breast-screening-helping-you-decide/nhs-breast-screening-helping-you-decide(gov.uk in Bing).
