National Screening Committee rejects use of AI in breast cancer screening
However, rather than completely writing off the potential of AI to massively improve the scope and efficiency of screening for breast cancer – the most common cancer in women in the UK – the report leaves the door open for further research into AI's accuracy in different settings, how well it performs for different ethnic groups and how much time it would save overloaded NHS staff.
Out of necessity, the report is extremely cautious, not least since in order to change the current screening programme, much more evidence is needed and that data are simply not available.
The study looked at how effective deep learning algorithms of AI were in examining breast cancer screening mammograms for signs of cancer. Currently, women aged 50 –70 years are invited for breast cancer screening every three years. Interpretation of each digital mammography is then conducted by two readers, both of whom decide whether the image appears normal or if a woman needs to be recalled for further tests.
A main reservation of the study was that AI could alter the spectrum of disease detected at breast screening and actually lead to an increase in over-diagnosis, meaning more women than should would be recalled, creating distress for them and also increasing the workload of NHS staff, rather than reducing it.
On the plus side, the potential of AI would be: one, it could replace one or all human readers; two, it could be used to pre-screen images with only high-risk images passed onto human readers; three, AI - which never suffers from over-work or fatigue - could be used as a reader aid, where the human reader uses the AI system for decision support.
I remain determinedly optimistic of the potential of AI to improve outcomes of testing, not least since I ran a case recently that involved misinterpretation of a mammogram, leading to very difficult outcome. The expert I instructed discussed the standards of interpretation of mammograms as generally being poor and was very positive about the potential benefits of AI in his work. Elsewhere in Europe, the European Union has approved a breast cancer screening system that combines a thermal imaging device with cloud-hosted AI analytics.
Clearly, anything that enables faster diagnosis and therefore treatment of breast cancer must be a good thing. This study concluded that further study into AI and breast cancer screening may be necessary in 1-3 years’ time and I very much hope for a different conclusion at that time.
Meanwhile, despite the NSC report, the Department of Health and Social Care has recently granted funding to study the accuracy of AI in breast cancer screening in the NHS to a research partnership that includes Imperial College London, Google Health, Imperial College Healthcare NHS Trust, St George’s Hospitals NHS Foundation Trust and the Royal Surrey NHS Foundation Trust. Researchers will examine how radiologists and clinicians interact with the AI system in a clinical setting.