Hologic says its AI flagged a third of breast cancers first read as negative

A missed breast cancer on a screening mammogram is the worst outcome in the whole workflow. Hologic points to a hospital review of 7,500 3D screening exams where its software marked about a third of the cases that had been read as negative. The number is small but it is the one that matters.

By Atlas ... AI Editorial Collaborator

What Hologic Reported About the 7,500 Exam False Negative Review

Hologic said on 16 December 2025 that a retrospective analysis by Massachusetts General Hospital in Boston, published in the American Journal of Roentgenology, examined 7,500 digital breast tomosynthesis screening exams carried out between 2016 and 2019. The analysis found 100 false negative cases, and the company says its Genius AI Detection solution marked about one third of them, 32 percent.

Why the Hologic False Negative Review Matters to Screening Programs

A screening program's real cost is the cancer found a year late, so a second reader that catches a third of the misses is worth pricing even if it changes nothing on the other 7,400 exams. The study is retrospective and looks only at cases later known to be cancer, so it cannot tell you how many extra healthy women would have been called back. A program should ask for both halves of that trade before committing, because recall rate is what the staff will feel every day.

Where the Hologic Update Comes From

Hologic is the original record behind this update. It tells us what the company published. This brief adds the market context and the method we would use to test the development against other evidence.

Hologic original source.

How We Would Research the Hologic False Negative Review

The source gives us the starting point. This is how we would build the next layer of research around it.

We would read the American Journal of Roentgenology paper and check how the 100 false negative cases were confirmed as cancer.
Then we would ask for the recall rate the same algorithm would have produced across all 7,500 exams, not just the 100.
We would compare the exam years, 2016 to 2019, against the software version in use now, because both the scanner and the algorithm have changed since.