AI system is better than human doctors at predicting breast cancer

An artificial intelligence system is better at predicting breast cancer than radiologists, according to a UK-US study led by Google Health. The team behind the technology hopes it can be widely deployed to improve cancer care.

Catching cancer early improves the chances of treatment succeeding. That is why many countries routinely screen women for signs of breast cancer using an X-ray scan called a mammogram.

In the UK, women aged between 50 and 71 are invited for a scan every three years. The American Cancer Society recommends annual scans for women aged between 45 and 54, and suggests that women aged 55 and over have a scan every one or two years.

But such programmes are far from perfect. Radiologists can miss signs of cancer in some women, while others may be prescribed harsh chemical and surgical treatments for lesions that might never have caused cancer. Over a 10-year period, half of women in the US who have a mammogram will have a false positive result. In 2014, one large Canadian study found that women given annual breast scans were no more likely to survive breast cancer in the long run.

Huge numbers

In an attempt to improve diagnoses, Shravya Shetty at Google Health and her colleagues trained an AI system on 91,000 mammograms taken from women in the UK and US. In each case, women were followed for two or three years to confirm whether or not they developed breast cancer. The team then tested their AI system on 28,000 other mammograms.

When compared with radiologists’ opinions, the AI system delivered 5.7 per cent fewer false positive results in the US sample and 1.2 per cent fewer false positives in the UK sample. Those figures may sound small, but given that 65 per cent of women aged 40 and over in the US have had a mammogram in the past two years, they represent huge numbers of women.

In another experiment, the team pitted its AI against six radiologists working individually. The AI performed 11.5 per cent better than the radiologists, even when the humans were given patient and family histories, says Shetty.

The system doesn’t do better than two radiologists working together – the number required to assess mammograms in the UK and other European countries – points out Adam Brentnall at Queen Mary University of London. “But the results appear promising,” he says.

The AI system could provide a second opinion, “without having the expense and difficulty of finding a second radiologist”, says Dominic King at Google Health.

The team, which has previously developed a similar technology for lung cancer screening, is investigating other uses for the technology.

Journal reference: Nature, DOI: 10.1038/s41586-019-1799