Artificial Intelligence improves breast cancer detection

Lisa Jarvis, Tribune News Service

A large, rigorous study in Sweden of artificial intelligence in breast cancer screenings suggests AI can help doctors detect cancers more efficiently. We need more such studies to determine when the technology has real value — and when it might have risks. And although the findings are incredibly promising, because Europe uses different processes and technologies for cancer screening, the US needs to commit to running its own similar studies to guide doctors here.

Previous large studies all looked back at old medical records to gauge whether AI was capable of detecting cancers as accurately as doctors. This study is the first trial of its size to test AI in real time on real patients — and will one day tell the field whether it actually improves the health of women. All of this is critical information as the technology increasingly becomes integrated into health care. In the study, some 80,000 women in Sweden were randomly assigned to either receive a double reading, where two independent radiologists look at the mammogram, or an AI-supported screening, which was performed by one radiologist and a computer.

The first stage of the study, the results of which were reported in the Lancet Oncology, was designed to ask whether it was safe to integrate AI into practice. The answer is a resounding yes. Overall, the computer helped humans flag more cancers, detecting about 20% more cancers than the two radiologists. Impressively, it did so with about the same rate as false positives (e.g. screens that looked like cancer but didn’t turn out to be).

Moreover, the researchers clearly showed that AI can reduce the workload for radiologists. Although the team didn’t directly measure the number of hours saved by using a computer to analyse mammograms, they estimate that the technology reduced screen reading time by about 44%. “In a situation where the medical workforce is strained, that’s a significant improvement,” says Larry Norton, medical director of the Evelyn H. Lauder Breast Center of Memorial Sloan Kettering Cancer Center. Even if the technology doesn’t turn out to be more accurate than doctors at picking up cancers, being just as accurate but faster would still be a major advance, he says.

Now comes the hard work of showing that this improves cancer care. “The holy grail is really understanding whether this kind of technology improves health,” says Yale School of Medicine’s Ilana Richman, whose research focuses on evaluating new breast cancer screening technologies. “We won’t know that for some time.”

The researchers in Sweden will continue to study the women in their trial to try to answer that question. In addition to confirming AI’s performance in detecting cancers, they will probe whether those additional cancers being detected are meaningful — that is, are the additional early lesions caught by the computer ones that would eventually cause a woman harm? They will also ask whether the method can reduce the number of “interval cancers,” or ones that are found between screenings and tend to be more aggressive and deadlier.

The need for this type of careful evaluation of AI is clear. So-called computer-aided detection that used more rudimentary versions of AI was widely adopted (particularly after Congress required Medicare to cover its use), but it led to an increase in false positives and biopsies for precancerous cells that are not typically dangerous. All of that came at a cost to the health-care system — when someone was flagged by a computer, they typically would go on to other types of tests and procedures that weren’t always needed. For now, any efficiencies that come out of the study will mostly benefit people in Europe and Australia, where breast cancer is typically screened by a team of two radiologists that might be safely reduced to one plus a computer. Translating the results to the US is complicated by the different standard of care — mammograms are reviewed by just one radiologist and are typically a 3-D scan rather than the 2-D ones used in the Sweden study.

But there are still some lessons for the US. For example, the algorithm used in the study was remarkably good at stratifying women by cancer risk — low, intermediate or high. And it turned out that the small number of women filtered into the high-risk group had a large portion of the cancers in the study. That points to the potential to use AI to triage the sea of exams passing before a radiologist each day, helping them prioritise the high-risk ones. That could lead to patients getting treated as quickly as possible, says Laura Heacock, a radiologist at NYU Langone Health. Patients might be wondering where all of this is leading: One day, will their cancer be diagnosed solely by a computer? That’s a prediction that Geoffrey Hinton, one of the so-called godfathers of AI, made back in 2016. “If you work as a radiologist, you’re like the coyote that’s already over the edge of the cliff but hasn’t yet looked down so doesn’t realise there’s no ground underneath him,” he said, suggesting they would be obsolete in five to 10 years. “People should stop training radiologists now.”