Invasive Breast Cancers Missed by AI Screening of Mammograms

Area of Science:

• Radiology
• Artificial Intelligence in Medicine
• Oncology

Background:

• Limited understanding of features of invasive breast cancers missed by AI on mammograms.
• Need to assess AI performance across different breast cancer molecular subtypes.

Purpose of the Study:

• To determine the false-negative rate (FNR) of AI mammogram evaluation based on molecular subtype.
• To identify characteristics and reasons for cancers missed by AI.

Main Methods:

• Retrospective analysis of 1097 mammograms from 1082 patients diagnosed with breast cancer (2014-2020).
• Commercial AI software used to evaluate mammograms; FNR calculated by molecular subtype (luminal, HER2-enriched, triple-negative).
• Radiologist review to classify AI-missed cancers and determine reasons for misses.

Main Results:

• AI missed 14% of cancers; FNR was lowest for HER2-enriched subtype (9%) compared to luminal (17.2%) and triple-negative (14.5%) (P=.001).
• AI-missed cancers were associated with younger age, smaller tumor size (≤2 cm), lower grade, fewer lymph node metastases, BI-RADS 4 findings, lower Ki-67, and nonmammary zone locations.
• Common reasons for AI misses included dense breasts (56), nonmammary zone locations (22), architectural distortions (12), and amorphous microcalcifications (5).

Conclusions:

• AI missed cancers were more frequent in luminal subtypes.
• Strategies to reduce AI misses should focus on dense breasts, nonmammary zone locations, architectural distortions, and amorphous calcifications.
• Targeted improvements are needed for AI performance in mammogram interpretation, particularly for specific cancer subtypes and imaging features.