Common features of breast cancers missed by AI
New research published Tuesday explores some of the common features of breast cancers missed by artificial intelligence.
AI software has been regarded as a promising tool for detecting early-stage breast cancer while also reducing radiologists’ workloads. However, previous research has shown this technology can sometimes overlook cases among certain women, experts explained in RSNA’s Radiology.
South Korean scientists recently investigated this phenomenon, examining a dataset of breast images from nearly 1,100 women screened between 2014 and 2020. Based on their results, they advocate for “meticulous assessments” of cases that include luminal cancer, dense breasts, and lesions located outside the mammary zone, among other factors.
“Although artificial intelligence is useful for detecting advanced-stage invasive cancers, it is inadequate for identifying cancers with some of the features revealed in this study,” lead author Ok Hee Woo, with the Department of Radiology at Korea University Guro Hospital, Seoul, and colleagues concluded. “Understanding the features of AI-missed invasive cancers on mammograms can help readers use AI appropriately in clinical practice, thus contributing to its further optimization.”
For the study, a breast radiologist with 14 years of experience reviewed the mammograms and classified lesions. Researchers utilized commercially available AI software from Seoul-based Lunit to read consecutive exams. “AI-missed cancers” were defined as those for which the software did not pinpoint a precise location matching the reference standard established by the expert radiologist. Three rads further assessed the exams with cancers overlooked by AI and determined whether they could have been acted upon, along with reasons for the mistake.
Altogether, artificial intelligence missed about 14% of cancers (or 154 of 1,097). AI-missed cancers were associated with patients of a younger age, a tumor size less than or equal to 2 centimeters, a lower histologic grade, fewer lymph node metastases, and more Breast Imaging Reporting and Data System category 4 findings, among other factors. In blinded reviews handled by radiologists, the reasons for misses were most commonly dense breasts (n = 56), nonmammary zone locations (22), architectural distortions (12), and amorphous microcalcifications (5).
In a corresponding editorial, radiologist Lisa A. Mullen, MD, noted the study has potential limitations, including its homogenous, all-Asian patient population and high percentage of patients with dense breasts (over 70%). However, she believes the findings further contribute to the specialty’s understanding of breast imaging AI, with key clues for physicians confronting similar scenarios.
“When using AI, it is critical for radiologists to understand what could be potentially missed by the software so that the radiologist can use the information to decrease the chance of missed cancers,” wrote Mullen, with the department of radiology at Johns Hopkins in Baltimore. “The radiologist should pay close attention to dense breasts and nonmammary zone areas, as well as search carefully for architectural distortion, microcalcifications, and small lesions. Areas of future research should include similar studies with more diverse patient populations and evaluation of other AI algorithms. Continued improvement of AI algorithms is also indicated to increase cancer detection and limit false-negative interpretation.”
Marty Stempniak has covered healthcare since 2012, with his byline appearing in the American Hospital Association's member magazine, Modern Healthcare and McKnight's. Prior to that, he wrote about village government and local business for his hometown newspaper in Oak Park, Illinois. He won a Peter Lisagor and Gold EXCEL awards in 2017 for his coverage of the opioid epidemic.