This AI tool is designed to work independently, dividing scans between those that need no radiologist assessment and those that require further interpretation.
Radiologists who use a commercially available artificial intelligence (AI) cancer detection tool to divide mammograms into scans that need radiologist assessment and those that do not could drop their workload by more than 50 percent, according to a newly published study.
Researchers from Sweden published results of their study that examines the AI tool as an independent reader on Sept. 1 in The Lancet Digital Health. That ability to assess diagnostic images without the active involvement of a radiologist sets this algorithm apart from other cancer-detection tools designed for mammography.
“Commercial AI algorithms as independent readers of screening mammography assessment are now performing on a clinically relevant level,” wrote the team led by Karin Dembrower, M.D., a breast radiologist at Capio Sankt Görans Hospital. “AI-based scoring can be used to reallocate radiologist time from clearly negative mammograms towards cases where cancer might go undetected. AI has the potential to promote early detection and, thereby, increase overall survival for breast cancer patients.”
Related Content: FDA Clears First AI Solution for 2D and 3D Mammography
In fact, according to the study analysis, the team said, using this AI software could potentially take more than 50 percent of these types of scans off of a radiologist’s worklist without any risk of missing cancers. And, in the cases of women who have the highest scores, the software could have picked up 27 percent of cancers.
To determine how well the algorithm worked, the team conducted a retrospective simulation study. They included 7,364 women between the ages of 40 and 47. Of the study population, 547 had a breast cancer diagnosis, and 6,817 were healthy.
Using these groups, Dembrower’s team evaluated two scenarios. For the first, the algorithm labeled a scan as normal with no need for further radiologist attention. Scans included in the second exam were deemed normal after two readings, but the algorithm identified them as high risk and tagged them as needing a follow-up MRI.
Based on their analysis, for the 60 percent, 70 percent, and 80 percent of women who had the lowest AI scores, the proportion of screen-detected cancers that could have been potentially missed were 0 percent, 0.3 percent, and 2.6 percent, respectively.
But, the story was different when they looked at the 1 percent or 5 percent of women who had the highest AI scores. In that group, 12 percent to 27 percent, respectively, of the 200 subsequent interval cancers would have been overlooked. In addition, that group would have had 14 percent to 35 percent, respectively, of 347 next-round screen-detected cancers go undetected.
While additional research is necessary to validate these findings, the team reiterated the beneficial impact this AI tool can have on radiologist workflow.
“Using a commercial AI cancer detector to triage mammograms into no radiologist assessment and enhanced assessment could potentially reduce radiologist workload by more than half,” the team said, “and pre-emptively detect a substantial proportion of cancers otherwise diagnosed later.”