Screening mammograms are key in diagnosing breast cancer early, but, according to the National Cancer Institute, they miss about 20% of breast cancers. There is hope that artificial intelligence could improve these figures, or at least help radiologists detect abnormalities faster. A new study investigated how much of a help AI can be in this regard.

A study recently published in the journal Radiology determined the false-negative rate of AI-assisted mammograms for several tumor subtypes. The researchers included 1.097 invasive breast cancer cases diagnosed between 2014 and 2020. AI software was used to read the associated mammograms, and the team found that it missed 14%, or 154, of these cases.

Success seemed to be linked with subtype, as 17.2% of luminal-type cancers were missed, while only 9% of HER2-positive cancers were not identified. For triple negative breast cancer, the error rate was 14.5%. In all, 61.7% of these tumors were considered detectable by radiologists.

Errors were most common in younger women and those with tumors two centimeters or smaller, a lower histologic grade, fewer lymph node metastases, lower Ki-67 expression, and located in non-glandular areas. Researchers say the main reasons behind the missed tumors were dense breasts, the fact that they were in non-glandular locations, structural distortions, and microcalcifications.  The findings suggest there needs to be more work on AI analysis of mammograms in these particular areas.

Professor Sung Eun Song, study co-author from the Department of Radiology at Korea University Anam Hospital, explains, “While AI shows strong performance in detecting breast cancer, our findings highlight the need for continuous oversight and complementary work by radiologists. Knowing the features of invasive cancers that AI tends to miss will be key to improving both clinical use and future development of AI tools.”

If you’d like to contribute to more life-changing research into breast cancer, click here!