Hologic says new data adds to the growing evidence supporting AI’s potential in identifying breast cancers that might be missed on screening mammography.

In a study published in the American Journal of Roentgenology, researchers at Massachusetts General Hospital in Boston analyzed 7,500 digital breast tomosynthesis screening exams conducted between 2016 and 2019 using Hologic’s Genius AI Detection software. The retrospective review focused particularly on false-negative cases; mammograms initially read as negative but later followed by a breast cancer diagnosis within a year.

“At Hologic, we are committed to continuous innovation to enhance the quality and reliability of our technologies,” said Mark Horvath, president of breast and skeletal health solutions at Hologic. “This study underscores AI’s potential to detect cancers that may otherwise remain undetected, providing us with valuable insights to guide future advancements.”
Among the 100 false-negative exams identified, the AI software flagged roughly one-third (32%) as suspicious and accurately localized the area where cancer was later diagnosed. In a separate analysis of 500 cancers previously detected by radiologists, the AI identified nearly 90% and pinpointed their locations with high accuracy.

The technology was particularly effective in flagging invasive ductal carcinomas and lymph node-positive cancers but less accurate with invasive lobular carcinomas and grade I invasive cancers. One case from the study stands out: a 54-year-old woman whose screening mammogram was initially interpreted as negative. Eleven months later, she returned with a palpable lump and was diagnosed with grade I invasive ductal carcinoma. On retrospective review, the AI algorithm had flagged and correctly localized the area where the cancer was later found.

Manisha Bahl, M.D., MPH, associate director of quality for breast imaging at Mass General Brigham and associate professor of radiology at Harvard Medical School, said the AI not only flagged the exam as suspicious but also correctly localized the area of concern — underscoring its potential value as a second set of eyes. Still, she emphasized that further research is needed to better understand how the technology could fit into clinical practice.

The authors also cited several limitations. For starters, the analysis was conducted at a single academic medical center with a predominantly Caucasian patient population and evaluated only one version of Hologic’s AI software. Also, the study did not assess patient outcomes or real-world clinical workflow performance, and some subgroup analyses were constrained by small sample sizes.

In a statement, Hologic said its FDA-cleared Genius AI Detection software has been trained and validated using a diverse, multisite data set. The company pointed to prior research involving more than 7,500 digital breast tomosynthesis exams across Asian, Black, Hispanic, and White patient cohorts, which found comparable algorithm performance across groups.

U.S. Healthcare Homepage