Breast Imaging AI: Catching Interval Cancers and Streamlining Screening

Computer-aided detection (CAD) is not new in breast imaging, but two new studies show that artificial intelligence (AI) can improve, not only the detection of interval cancers, but it can also be used to alleviate the workload burden that is attached to many screening programs.

During a Nov. 30 session of this year’s Radiological Society of North America (RSNA) annual meeting, researchers from the United Kingdom outlined how putting these tools in place can streamline and improve cancer detection.

Improving Interval Cancer Detection

Although the number of interval cancers are small, said Sarah Jane Vinnicombe, M.D., a consultant radiologist at Thirlestaine Breast Center in Cheltenham, U.K., identifying them can still be critical to good patient care.

Based on her study that retrospectively applied an AI algorithm to screening mammograms, she found that the tool can successfully pinpoint these malignancies.

“AI-based CAD does appear to be able to localize some interval cancers on prior screening that are missed by expert readers,” she said, “and, the cancers detected are mostly in the good prognosis group.”

To reach these results, her team reviewed screening mammograms collected between 2011 and 2016, 300 of which had interval cancers and 853 that were controls. Of the exams, 42 percent of exams with interval cancers had high breast density while only 36 percent of the control group did.

Based on the analysis, there were 109 minimal sign, false negative scans. In that group, the AI tool flagged 21 lesions (19 percent) – 19 were determined to be invasive. Of those, 15 lesions – 79 percent – were either grade 1 or 2, and all were ER-positive. At 90-percent specificity, though, the tool pinpointed 42 lesions (39 percent), with 38 being invasive. Seventy-one percent were grades 1 or 2, and 26 were ER-positive.

In addition, she said, CAD yielded high global scores 21 percent of true interval cancers at 90 percent specificity.

“This is a group that might require enhanced screening,” she said.

AI for Screening Program Second Reads

Most screening programs require double-reading to determine whether a woman should be referred for follow-up imaging, tying up the time of two radiologists. But, according to a study presented by Nisha Sharma, M.D., a breast imaging consultant with the National Health Service in the United Kingdom, replacing one radiologist with AI can catch the same number of cancers while freeing up a provider for other responsibilities.

“[Using AI] creates a feasible solution to combat the workforce crisis within breast imaging,” she said. “This solution would allow for the opportunity to create efficiency in the workforce where radiologists could be redeployed to manage clinics rather than be required to film read.”

To reach this conclusion, she said, the team reviewed 40,588 mammograms that were initially read by two human readers, 40,230 of which were normal and 358 of which had biopsy-proven cancers, resulting in a 4-percent recall rate. Roughly 3 percent required a third review due to disagreement between readers.

When using the algorithm, the number of exams needing a third review grew to 18.1 percent, and, when used alone its sensitivity was 85.5 percent and its specificity was 87.2 percent. This fell below the 89.4 percent and 96 percent, respectively, produced by first human readers.

The benefit of AI emerged when it was paired with a first human reader, though, she said. At that point, its sensitivity and specificity rose to 95 percent and 96.9 percent, respectively. The disagreement rate 8.4 per 1,000 scans – similar to that of two human readers – and the recall rate was, again, 4 percent.

Ultimately, she said, the study underscores that replacing a second human reader with AI is a viable option for screening mammography programs.

For additional RSNA coverage, click here.