New research on 175,000 women – the largest NHS study to date – on the use of AI in breast cancer screening shows that AI detected more cases of invasive cancer, more cases overall, had fewer false positives, and recalled fewer women having their first scan than humans did. For one part of the study, AI reduced the time spent reading scans by almost a third.
The research was conducted by Imperial College London, Google, the universities of Cambridge and Surrey, NHS Trusts at Cambridge University Hospitals, Imperial College Healthcare, the Royal Marsden, the Royal Surrey and St George’s University Hospitals, and the AIMS public engagement group.
One case every 10 minutes
Breast cancer is the most common cancer in the UK, with one woman diagnosed every 10 minutes.
Yet there is a 29 per cent shortfall of clinical radiologists – almost 2,000 – and this is predicted to rise to 39 per cent by 2029. This research demonstrates the potential for more women to be diagnosed and treated sooner, while reducing radiologists’ workload.
In the UK, breast cancer scans are assessed by two readers, usually specialist radiologists. Each reads the scan separately, with the second reader either knowing or not knowing the first reader’s decision.
This research looked at how two human readers compared to one human reader plus one AI reader, using AI software developed by Google. The findings are published in two linked papers out today in Nature Cancer.
There were three parts to the study.
The first was a retrospective study of 125,000 women, aged 50 to 70, from five NHS screening services. They were screened in 2015-16 over a 39-month follow-up period. The final analysis included 115,973 breast cancer scans.
AI as the second reader achieved better results than the first human reader, with the cancer detection rate (CDR) rising from 7.54 (human) to 9.33 (AI) per 1,000 women.
AI also identified more invasive cancers, significantly reduced false positives, and detected 25 per cent of interval cancers (cancers detected between healthy scans).
AI performed particularly well for first screens – with 39.3 per cent fewer recalls and a 8.8 per cent higher CDR.
With AI, the amount of time taken to read a scan reduced by almost a third (32.1 per cent or 195,983 vs 288,616 reads), representing a significant reduction in workload.
The second part of the study looked at 9,266 current cases at two screening services at 12 sites in London.
During the first two weeks, AI had a higher recall rate than the humans did and was above the target recall rate for the study, so the researchers adjusted the criteria. Despite this, AI continued to have a higher recall rate.
Big time savings
However, across both sites, the average time for AI to complete a read was 17.7 mins compared to 2.08 days for the first human reader, a significant time saving.
First use of AI in arbitration
The third part of the study looked at the use of AI in arbitration in 50,000 women.
Arbitration is when the first and second readers don’t agree on the diagnosis and a third reader analyses the scan and makes a final decision. This is the first time AI has been used in this scenario. It found that AI fared comparably to the humans.
AI had a higher arbitration rate but, on balance, reduced the overall screening workload. The researchers suggest that further development of the AI tool could potentially lead to the detection of cancers earlier than with two human readers.