Emerging research suggests that dense breasts, non-mammary zone locations and tumor sizes < 2 cm are key factors contributing to actionable false-negative cases with artificial intelligence (AI) mammography software.
For the retrospective study, recently published in Radiology, researchers assessed false negatives in AI assessment (Lunit Insight MMG, version 1.1.7.3, Lunit) of mammograms for 1,082 women (mean age of 54.3) with a total of 1,097 invasive breast cancers.
The study authors found that the AI software missed 154 out of 1,097 invasive breast cancers (14 percent) and that 61.7 percent of these cases were actionable.
Of the actionable AI-missed breast cancers, the researchers noted that 59 percent of these cases involved women with dense breasts and 23 percent involved non-mammary zone locations.
In contrast to patients with AI-detected cancers, the researchers found that women with AI-missed invasive breast cancers had over a 20 percent higher incidence of tumor sizes > 2 cm (81.8 percent vs. 61 percent); were over five years younger on average (mean of 49.7 years of age vs. 55.1 years of age); and had a nearly 13 percent lower incidence of axillary lymph node metastases (18.2 percent vs. 31.1 percent).
“Compared with AI-detected cancers, AI-missed cancers were associated with younger age, a tumor size less than or equal to 2 cm, a lower histologic grade, fewer lymph node metastasis, low Ki-67 expression, fewer HER2-positive tumors, luminal subtype, more BI-RADS category 4 interpretations, and frequent non-mammary zone locations,” wrote lead study author Ok Hee Woo, M.D., who is affiliated with the Department of Radiology at Korea University Guro Hospital in Seoul, Korea, and colleagues.
Three Key Takeaways
- AI misses are often in dense breasts and non-mammary zones.
AI-missed breast cancers were frequently associated with dense breast tissue (59 percent) and lesions located in non-mammary zones (23 percent), highlighting known challenges for mammographic interpretation. - Tumor and patient characteristics differ in missed cases.
Compared to AI-detected cancers, AI-missed cancers were more common in younger women, had smaller tumor sizes (≤ 2 cm), lower histologic grade, and fewer lymph node metastases, features that may contribute to lower AI detection sensitivity. - False-negative rates vary by molecular subtype.
The AI system had higher false-negative rates for luminal (17.2 percent) and triple-negative (14.5 percent) subtypes, and the lowest rate for HER2-enriched cancers (9 percent), likely due to higher detection of microcalcifications associated with HER2-positive tumors.
With respect to molecular subtypes of breast cancer, the researchers found higher false-negative rates (FNRs) with luminal and triple-negative subtypes (17.2 percent and 14.5 percent, respectively) in comparison to human epidermal growth factor receptor 2 (HER2) breast cancer (9 percent).
“Our finding that the HER2-enriched subtype had the lowest FNR and the highest AS (abnormality score) aligns with previous study results, showing AI’s highest detection rates and AS for this subtype due to the presence of microcalcifications,” added Woo and colleagues.
(Editor’s note: For related content, see “Mammography AI Platform for Five-Year Breast Cancer Risk Prediction Gets FDA De Novo Authorization,” “Mammography Study Reveals Over Sixfold Higher Risk of Advanced Cancer Presentation with Symptom-Detected Cancers” and “Emerging AI Mammography Model May Enhance Clarity for Initial BI-RADS 3 and 4 Classifications.”)
In regard to study limitations, the authors noted the retrospective design, the use of a single AI software and limiting the cohort to those with confirmed invasive breast cancer. They also acknowledged higher percentages of BI-RADS category 5 presentations (> 40 percent) and women with dense breasts (> 70 percent) that are not representative of broader populations.
