AI Detects Missed Interval Breast Cancer on Mammograms

TOPLINE:

An artificial intelligence (AI) system flagged high-risk areas on mammograms for potentially missed interval breast cancers (IBCs), which radiologists had also retrospectively identified as abnormal. Moreover, the AI detected a substantial number of IBCs that manual review had overlooked.

METHODOLOGY:

• Researchers conducted a retrospective analysis of 119 IBC screening mammograms of women (mean age, 57.3 years) with a high breast density (Breast Imaging Reporting and Data System [BI-RADS] c/d, 63.0%) using data retrieved from Cancer Registries of Eastern Switzerland and Grisons-Glarus databases.
• A recorded tumour was classified as IBC when an invasive or in situ BC was diagnosed within 24 months after a normal screening mammogram.
• Three radiologists retrospectively assessed the mammograms for visible signs of BC, which were then classified as either potentially missed IBCs or IBCs without retrospective abnormalities on the basis of consensus conference recommendations of radiologists.
• An AI system generated two scores (a scale of 0 to 100): a case score reflecting the likelihood that the mammogram currently harbours cancer and a risk score estimating the probability of a BC diagnosis within 2 years.

TAKEAWAY:

• Radiologists classified 68.9% of IBCs as those having no retrospective abnormalities and assigned significantly higher BI-RADS scores to the remaining 31.1% of potentially missed IBCs (P < .05).
• Potentially missed IBCs received significantly higher AI case scores (mean, 54.1 vs 23.1; P < .05) and were assigned to a higher risk category (48.7% vs 14.6%; P < .05) than IBCs without retrospective abnormalities.
• Of all IBC cases, 46.2% received an AI case score > 25, 25.2% scored > 50, and 13.4% scored > 75.
• Potentially missed IBCs scored widely between low and high risk and case scores, whereas IBCs without retrospective abnormalities scored low case and risk scores. Specifically, 73.0% of potentially missed IBCs vs 34.1% of IBCs without retrospective abnormalities had case scores > 25, 51.4% vs 13.4% had case scores > 50, and 29.7% vs 6.1% had case scores > 75.

IN PRACTICE:

“Our research highlights that an AI system can identify BC signs in relevant portions of IBC screening mammograms and thus potentially reduce the number of IBCs in an MSP [mammography screening program] that currently does not utilize an AI system,” the authors of the study concluded, adding that “it can identify some IBCs that are not visible to humans (IBCs without retrospective abnormalities).”

SOURCE:

This study was led by Jonas Subelack, Chair of Health Economics, Policy and Management, School of Medicine, University of St. Gallen, St. Gallen, Switzerland. It was published online on August 04, 2025, in European Radiology.

LIMITATIONS:

The retrospective study design inherently limited causal conclusions. Without access to diagnostic mammograms or the detailed position of BC, researchers could not evaluate whether AI-marked lesions corresponded to later detected BCs.

DISCLOSURES:

This research was funded by the Cancer League of Eastern Switzerland. One author reported receiving consulting and speaker fees from iCAD.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.