Artificial intelligence (AI) is set to slash the time it takes for women to receive a breast cancer diagnosis in Scotland, according to a senior clinician.
Studies carried out on women in Aberdeen have shown that the technology can also pick up more cancers, leading to a potential revolution in breast cancer screening.
Research trials looked at more than 80,000 mammogram scans obtained from women over a four-year period, using an AI tool and a “human reader”.
Results showed that the technology assisting a radiologist led to better detection rates and could even help predict future cancers in women, allowing treatment to be targeted earlier.
Clinicians are now working on ways of improving the computer algorithm further to ensure that it doesn’t recall too many scans for further checks when they are not necessary.
If successful, the new project, called Gemini, could usher in the use of AI in breast cancer screening programmes as standard within the next five years in Scotland.
Dr Gerald Lip, clinical director of the North East of Scotland Breast Screening programme, and who led the research, said: “I think this will impact on every one of the 200,000 women a year that we screen in Scotland.
“First, we will be able to show that the AI is as good as radiologists, so we don’t expect the cancer detection rate to drop.
“We might even expect it to go up by between another 10 to 20 per cent, reducing interval cancers, which are cancers that happen between screening rounds.
“Secondly, we will be able to reduce the turnaround time. So, on average, in some places where it takes up to 14 days, you could get the results almost within three days, reducing anxiety for the women involved.”
He added: “This is the first time ever that mammography AI has shown its effectiveness on Scottish women.”
Mammograms are currently looked at by two specialists, according to strict clinical safety and quality assurance guidelines. Dr Lip said that the aim is to replace one of the human readers in that process with an AI tool, which could reduce the workforce burden by up to 40 per cent.
Crucially, it would also respect the wishes of women involved in the study, who did not want their scans only to be analysed by a computer, according to a survey conducted of the participants.
Dr Lip said: “We did some engagement with the women attending the screening centre and they were positive about the role of AI in screening. But the key message that came through was that they wanted a human in the loop somewhere.
“And actually, because we have two humans that read the mammogram, you could look at the substituting of one of those humans with an AI and still maintain that high level of performance. Any disagreements between the technology and the human would be looked at by an extra person.”
But he did not rule out the possibility that AI could be used on its own in breast screening programmes if the technology approaches 100 per cent effectiveness.
“If we reach a scenario where we are almost 99.99 per cent sure that their mammogram is normal, then we may not need a human in the loop, but it’s a stage-by-stage process. We’re not going to get there straight away. We’re going to have show lots of levels of safety and monitoring.”
Dr Lip warned that like with any software package, there is always the chance that a new update could interfere with the functioning of the AI. He said constant vigilance will be needed over what is termed “drift” in the algorithm’s accuracy, which naturally occurs over time with AI tools as the data gets older.
“Anytime you bring something new into the system, you need to have your human fallbacks that you can switch back to,” he said. “There needs to be a very robust monitoring system in place for the AI under the medical device regulations to ensure that it is performing as it should.”
The study was funded by an Innovate UK grant and was supported by Scotland’s Industrial Centre for Artificial Intelligence Research in Digital Diagnostics (iCAIRD). The technology itself was provided by Kheiron Medical Technologies, a cancer diagnostics tech company, which developed the Mammography Intelligent Assessment tool (MIA).
The research focused on women attending a regional breast screening test centre in Aberdeen – which 25,000 women pass through every year – but the eventual aim is to introduce the AI system nationally in Scotland.
That would be relatively simple to rollout as the national breast cancer screening programme is already completely digital, unlike in England, Dr Lip said.
In numbers terms, the study published its anonymised data in November, looking at a “validation set” of data of 45,444 women, with 303 cancers. It found that the AI was able to detect 6.1 cancers out of 1,000 compared to a human reader of the scans at 5.38 out of 1,000, so 277 cancers compared to 261.
The number of scans “recalled” by the AI for further tests was 12.97 per cent compared to 5.74 per cent for a human. Dr Lip said the ideal “sweet spot” for the recall rate should be between four and seven per cent. “If we call too many women back, they get very worried and they won’t have anything. And if you call back too few, you’re not picking up the cancers you should,” he said.
The study indicated that the higher recall rate of the AI, however, may be beneficial, even when scans have been read as normal by a human reader.
Dr Lip said: “We were able to detect 30 per cent of the interval cancers that occurred, even though they were called normal by two humans in the original mammogram. So there seems to be some element of prediction, or maybe safety netting that it can pick up some of these additional cancers.
“So that led to the study that we’ve started now, Gemini, which looks at current mammograms, as they are being taken.” In that project – which launches soon – the AI will flag up an alert if it detects something of concern which a human reader has categorised as normal.
One of the main motivations for using AI in breast cancer screening is to support an ageing and under pressure radiologist workforce.
A census conducted by the Royal College of Radiologists in 2021 showed that the NHS radiologist workforce across the UK was short-staffed by a third and needs at least another 1,939 consultants to meet safe staffing levels and pre-coronavirus levels of demand for scans.
Dr Lip said: “The tiredness of colleagues came into much sharper relief during Covid, where we were all working really hard. A lot of people were choosing to retire as well.
“And we’re still struggling to catch up with the volume of work. If you look at the workload and the demand for imaging and radiology, it’s just going one way.”
He added: “It is my hope that radiologists will feel that they’re better supported, because it’s almost like having a friend or a second opinion with you.”
