A retrospective study of an artificial intelligence program for reading 3D mammography scans illustrated the pros, cons and limits of the technology, with researchers saying their findings could potentially help optimize the imaging algorithms used in breast cancer detection and diagnosis.
Conducted at Massachusetts General Hospital, the researchers used Hologic’s Genius AI Detection 2.0 mammography software to analyze a collection of 5,000 digital breast tomosynthesis exams performed between 2016 and 2019.
That dataset included 500 cases of cancer that were identified by radiologists at the time, as well as 100 patients with malignant tumors that had slipped by undetected.
The AI program correctly spotted about 90% of the known, true-positive cases, as well as 32% of the previously missed, false-negative cancers. According to the researchers, the AI was more likely to detect certain types of tumors and their presentations compared to others. The results were presented at the annual symposium of the Society of Breast Imaging, being held this week in Colorado.
Among the true-positive cases, the AI was less likely to identify cancers that looked like otherwise benign asymmetries between the two breasts. It was also less likely to spot invasive lobular carcinomas and grade I invasive carcinomas. However, it was more accurate in detecting invasive ductal carcinomas.
The software also tended to ping more true-positive cancers if the breast tissue was dense or if they presented as masses on the scan, though these findings were not statistically significant in the study. The AI was also more likely to uncover grade 3 invasive carcinomas, as well as lymph node-positive cancers.
When it came to the false-negative reports, the AI was more likely to detect lesions correctly if the radiologist at the time also identified a mammographic finding during screening—specifically, the presence of breast calcifications, which can be an early sign of cancer.
Meanwhile, out of the 4,400 exams determined to be cancer-free scans, the AI correctly categorized 55% of those cases as negative.
“As AI continues to evolve, I believe it will become an increasingly vital tool for radiologists, helping to transform breast cancer detection and ultimately reduce the burden of this disease for patients,” the study’s presenter, Manisha Bahl, said in a statement. Bahl is the director of the MGH’s breast imaging fellowship program and an associate professor of radiology at Harvard Medical School.
Hologic has said its updated Genius AI Detection PRO workflow platform—with added color coding and automated reporting—has been shown to reduce radiologists’ total reading time by 24%.
The Genius AI Detection 2.0 program was launched in 2023, with the company saying it reduced false-positive markings per case by more than 70% compared to its previous ImageChecker CAD solution.