• Posted August 2, 2023

AI-Assisted Mammograms Could Be a Big Advance: Study

Artificial intelligence (AI) programs can safely be used to help radiologists review mammogram images and detect breast cancers, early results from an ongoing clinical trial show.

A single radiologist aided by AI wound up detecting about 20% more breast cancers from mammogram images than two radiologists working together, according to a report in the August issue of The Lancet Oncology.

The use of AI did not increase the rate of false positives, where a mammogram is incorrectly diagnosed as abnormal and in need of follow-up, the researchers found.

At the same time, AI assistance reduced radiologists' workload by about 44%, the results showed.

“Our study showed that mammography screening with AI is safe, since the cancer detection did not decline despite the reduced screen-reading volume for the radiologist,” said lead researcher Dr. Kristina Lång, an associate professor of diagnostic radiology at Lund University in Sweden. “Women should, therefore, feel safe undergoing AI-supported screening.”

AI-aided mammography already is available in the United States, with more than 20 AI systems approved by the U.S. Food and Drug Administration, said Robert Smith, senior vice president of early cancer detection science with the American Cancer Society.

“No question it is the way of the future, but I have to emphasize that there's still a lot of questions about how we use it and how we determine how systems must perform in order to be used in the clinical setting,” he said.

Most studies of AI-aided mammography have been retrospective, with researchers testing the ability of computers to accurately evaluate past cases where doctors already know whether or not there's breast cancer, Smith said.

This new clinical trial, conducted in Sweden, is the first study to directly compare AI-assisted mammography against standard mammography.

More than 80,000 already have been randomly assigned to undergo either standard or AI-assisted mammography at four testing sites in Sweden, the researchers said. By the end of the trial, 100,000 participants will be enrolled and tracked for two years of follow-up.

The analysis is an interim safety report, to make sure that the women undergoing AI-assisted mammography aren't facing undue risk.

There's a shortage of breast radiologists in many countries, and accurate AI programs would relieve the workload, the study authors said in background notes.

The ultimate goal of the clinical trial is to see whether AI can help reduce the number of breast cancers missed during a screening. At this point, radiologists overlook highly suspicious early signs of malignancy in about 20% to 30% of breast cancers that are eventually detected, the researchers said.

For this interim report, they set a minimum limit for clinical safety at a detection rate of 3 cancers for every 1,000 screenings.

AI-supported screening resulted in a detection rate of 6 cancers per 1,000, compared with 5 per 1,000 for standard screening. That's the equivalent of detecting 1 additional cancer for every 1,000 women screened, according to the report.

Women were recalled for further testing in about 2.2% of cases involving AI-aided screening and 2% of cases with standard screening. That resulted in 41 more cancers detected with the support of AI, of which 19 were invasive.

The false-positive rate was 1.5% in both groups, the researchers said.

At the same time, there were nearly 37,000 fewer screen readings by radiologists in the AI-supported group.

At that rate, AI helped cut four to six months off the workload of radiologists assigned to read the 40,000 images taken at this point, assuming a radiologist reads an average 50 mammograms each hour, the investigators estimated.

“We found that AI-supported screening led to more detected cancers despite a substantial reduction of the screen-reading workload, without affecting false positives,” Lång said.

However, she noted that the study emphasized the central role of radiologists in reviewing mammograms, even with the help of AI.

“Currently, I do not believe that AI should operate alone," Lång said. "Human intelligence and AI complement each other, which I believe our study has illustrated.”

She said there are potential downsides to widespread use of AI in mammography that need to be considered.

“Radiologists might become over-reliant on AI over time, which could increase false positives, a negative effect of screening,” Lång said.

Radiologists also might find themselves overworked anyway, if the use of AI results in the redistribution of more complex tasks to humans while the computers handle the relatively simple problems, she added.

Final results from the trial will show whether AI can help catch cancers earlier, at a stage when they are more easily treated, the researchers noted.

“Is the AI picking up, for example, some slightly less aggressive tumors and possibly missing some more aggressive tumors?” Smith said. “We can't compare these numbers right now without digging deeper, although it's a good sign that they picked up 20% more cancers. That's encouraging.”

Lång hopes other similar trials will be conducted as follow-ups.

“The consequences of using AI need to be understood before a large-scale implementation should be considered,” she said. “Our results show great potential of the technique, but one study is not enough. Other prospective studies are needed to see whether our findings hold true for other screening programs and for other populations.”

More information

The U.S. National Cancer Institute has more about mammograms.

SOURCES: Kristina Lång, MD, PhD, associate professor, diagnostic radiology, Lund University, Lund, Sweden; Robert Smith, PhD, senior vice president, early cancer detection science, American Cancer Society, Atlanta; The Lancet Oncology, August 2023