What New Research Reveals About AI Software for Breast Ultrasound

Can AI enhance specificity for cancer detection with breast ultrasound?

In a retrospective study, recently presented at the Society of Breast Imaging (SBI) Symposium, researchers evaluated the use of adjunctive AI software (Koios Medical) for breast ultrasound in 526 women (total of 600 breast lesions).¹

The study authors found an 86 percent specificity rate with the adjunctive AI software in comparison to 70 percent for unassisted radiologist interpretation. The researchers also noted that the AI software downgraded 43 percent of cases that were previously interpreted by radiologists as BI-RADS 3 or higher lesions with two of these cases being cancers.¹

“These findings, which are encouraging, suggest that AI may help reduce unnecessary short interval follow-up imaging and benign biopsies, both of which are common challenges in breast imaging, at the same time,” noted study author Manisha Bahl, MD, MPH, FSBI, in a recent interview with Diagnostic Imaging.

However, Dr. Bahl pointed out that fellowship-trained radiologists provided a 99 percent sensitivity rate in contrast to 91 percent with the use of AI.¹

“These findings suggest that although AI may improve specificity, there may be a trade-off in missed cancers if the technology is relied upon too heavily,” explained Dr. Bahl, an associate professor of radiology at Harvard Medical School.

Calling for prospective real-world studies to provide more clarity on the utility of the AI software, Dr. Bahl said future research should examine patient-centered outcomes, such as potential reduction in benign biopsy rates, as well as the impact for radiologists of different experience levels.

Reference

1. Bahl M. Assessing the performance of an AI decision support tool for diagnostic breast ultrasound interpretation. Presented at the Society of Breast Imaging Symposium, April 16-19, 2026, Seattle. https://2026-sbi-symposium.eventscribe.net/index.asp