Artificial Intelligence (AI) systems are as accurate as human radiologists at detecting breast cancer in mammography screenings, and sometimes more accurate in crucial early detection of tumours, medical researchers in the Netherlands have found.
The study by scientists at Radboud University Medical Centre, published in the latest edition of The Lancet Digital Health, showed AI has the potential to reduce mortality by supporting breast cancer screening programmes often struggling with inadequate resources and staffing.
Current practice in the Netherlands requires two radiologists to read a mammogram. The Radboud researchers have found, however, that replacing the second radiologist with AI leads to results that are just as accurate.
The real strength of the AI systems, the team said, was in their ability to sound early warnings by identifying tumours human radiologists may have missed.
RM Block
The researchers analysed 42,000 breast scans in the Utrecht region routinely performed as part of the Dutch national screening programme. They followed the women who provided the scans for the next four-and-a-half years.
That research provided invaluable context for the study. “For example, sometimes AI sees a tumour but the radiologists don’t recognise it as such, and categorise it as a false positive”, said Suzanne van Winkel, one of the study’s authors.
“However, in our study, when the abnormality resurfaced in a subsequent scan and was then accurately identified by the radiologist, we knew the AI had been right first time.”
This gradual building of “trust” in the AI is likely to lead in the medium term to situations where the two-radiologist requirement is dropped and replaced by a combination of one radiologist and the AI system.
“Only if the AI has any doubts will a second radiologist be consulted.”
[ New breast cancer diagnosis technology aims to cut waiting timesOpens in new window ]
This would dramatically reduce the workload of radiology departments and save hospitals tens of millions of euro annually, according to team leader, Dr Ritse Mann of Radboud and the Netherlands Cancer Institute in Amsterdam, who describes the study as “a major outcome”.
In the short term, though, more investment in IT is essential to implement the project.
“If we work hard, the Netherlands is still about two years away from replacing the second radiologist with AI, but in reality, that transition will probably take four or five years,” said Dr Mann.
“In theory, however, with adequate investment, given what we’ve learned from this study, everything could be done in that time frame by autonomous AI – literally everything unless there’s an abnormality.”
Even so, AI won’t make radiologists obsolete, he said.
“It will, for instance, free up time for doctors to talk to their patients and make diagnoses instead of reading mammograms, which computers have now shown they can learn to do. It will certainly change the profession.”
