Hologic has put forward more data bolstering the use of its artificial intelligence programs for detecting breast cancer through mammography.
The study, conducted at Massachusetts General Hospital and recently published in the American Journal of Roentgenology, reviewed 7,500 digital 3D mammograms using Hologic’s Genius AI Detection platform.
Researchers found that between 2016 and 2019 there were 100 false-negative cases. Hologic’s platform flagged roughly one-third of those cases as having suspicious areas, then accurately identified the locations where breast cancers were later confirmed.
Of the 500 cases radiologists had previously found, the AI picked up 90% of them as well as correctly localizing their placements, according to the company.
“This study underscores AI’s potential to uncover cancers that might otherwise remain hidden, while also giving us critical insights to guide the development of future innovations,” Mark Horvath, president of Hologic’s breast and skeletal health division, said in a statement. “As we continue to advance this technology based on customer and provider feedback, we’re excited to see its impact in real-world settings.”
The study also found Genius AI tech was more likely to flag invasive ductal carcinomas and lymph-node-positive cancers in the study, though it was less likely to flag invasive lobular carcinomas and grade I invasive carcinomas, the company said.
Of the false-negative cases uncovered in the study, one focused on a 54-year-old woman who, after getting a negative result, returned to her doctor 11 months after she noticed a lump in her left breast. She was then diagnosed with grade 1 invasive carcinoma. The Hologic study’s retrospective review of the initial screening marked and correctly localized those areas as suspicious.
The latest data come just shy of a month after Hologic released a pair of separate studies supporting its mammography AI. Those studies found the use of AI tools can maintain high breast cancer detection rates while also streamlining radiologists' workflows and scan-reading times.