AI-Backed Mammography Keeps Up With Docs for Cancer Detection

Artificial intelligence-supported mammography resulted in a cancer detection rate comparable to that of two breast radiologists working together, while also substantially cutting screening workload.

Among more than 80,000 women randomized to either AI-supported mammography screening or to standard double reading without AI, cancer detection rates were 6.1 (95% CI 5.4-6.9) per 1,000 screened participants in the intervention group, which was above the lowest acceptable limit for safety, and 5.1 (95% CI 4.4-5.8) per 1,000 in the control group for a 1.2 ratio (95% CI 1.0-1.5, P=0.052), according to Kristina Lång, MD, PhD, of Lund University in Malmo, Sweden, and colleagues.

They reported in Lancet Oncology that AI-supported screening detected one additional cancer for every 1,000 women screened. They also observed that the use of AI did not influence the rates of recalls, false positives, or consensus meetings.

MASAI is the first randomized trial investigating AI in a national breast cancer screening program, the researchers said.

"AI-supported screen-reading procedure can be considered safe," they stated, noting that the benefit in terms of screen-reading workload reduction "was considerable."

Specifically, there were 36,886 fewer screen readings by radiologists in the AI-supported group than in the control group (46,345 vs 83,231), translating into a 44% reduction in the screen-reading workload of radiologists.

"The greatest potential of AI right now is that it could allow radiologists to be less burdened by the excessive amount of reading," said Lång in a news release. "While our AI-supported screening system requires at least one radiologist in charge of detection, it could potentially do away with the need for double reading of the majority of mammograms, easing the pressure on workloads and enabling radiologists to focus on more advanced diagnostics while shortening waiting times for patients."

Lång and colleagues reported that recall rates were 2.2% (95% CI 2.0-2.3) in the intervention group and 2.0% (95% CI 1.9-2.2) in the control group, while the positive predictive value of recall was 28.3% (95% CI 25.3-31.5) and 24.8% (95% CI 21.9-28.0), respectively. The false-positive rate was 1.5% (95% CI 1.4-1.7) in both groups.

In the intervention group, 75% (n=184/244) of cancers detected were invasive and 25% (n=60) were in situ; in the control group, those data came in at 81% (n=165/203) and 19% (n=38), respectively.

Lång and colleagues pointed out that there has been concern AI could lead to an increase in resource-demanding consensus meetings, but found that the proportion of screening that resulted in these meetings was not affected by the use of AI.

In an accompanying comment, Nereo Segnan, MD, MSc, and Antonio Ponti, MD, MPH, both of the Piedmont Reference Center for Epidemiology and Cancer Prevention in Turin, Italy, said that the outcomes seemed straightforward in favor of AI-assisted screening, but still advised caution when interpreting the results. They called for considering "the possible presence of overdiagnosis ... or overdetection of indolent lesions such as a relevant portion of ductal carcinomas in situ."

In the prespecified clinical safety analysis of the trial, AI identified a higher percentage of in situ carcinoma among screen-detected cancers, they noted. "There have been decades of debate on whether detection of in situ carcinomas, especially those classified as low grade, is beneficial or harmful in breast cancer screening," Segnan and Ponti said, adding that it is important to acquire biological information on the detected lesions.

They said an "important research question thus remains: is AI, when appropriately trained, able to capture relevant biological features -- or, in other words, the natural history of the disease -- such as the capacity of tumors to grow and disseminate?"

MASAI was done from April 2021 through July 2022 with 80,033 women (median age 54) who had undergone mammogram screening at four sites in southwest Sweden. Participants could have a moderate hereditary risk of breast cancer and/or a history of breast cancer. They were randomly assigned 1:1 to standard analysis done by two radiologists without AI, or to AI-supported screening in which an AI-supported reading system assessed mammograms before they were read by one or two radiologists.

Screen reading and additional assessment of recalled participants were done at a single site, the authors said, and that was a study limitation. Also, the combination of one type of mammography device and one AI system was used.

The researchers explained that their "screening strategy emphasizes the central role of the radiologist to make the final decision to recall a patient, and the present results are dependent on the performance of the participating radiologists."

The trial's primary outcome measure is interval cancer rate, which will be assessed after the full study population of 100,000 screened participants have had at least a 2-year follow-up.

Comment