Mammographic classification of interval breast cancers and artificial intelligence performance

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• European studies indicate artificial intelligence (AI) can reduce interval breast cancers (IBCs).
• Limited research exists on IBC classification and AI effectiveness in the U.S., especially with digital breast tomosynthesis (DBT) and annual screening.
• This study aimed to classify IBCs mammographically and evaluate AI performance within a 12-month screening interval.

Purpose of the Study:

• To classify interval breast cancers (IBCs) identified through mammography.
• To assess the effectiveness of a deep-learning AI tool in flagging and localizing IBCs.
• To analyze AI performance in relation to different IBC classifications and screening modalities (DM, DBT).

Main Methods:

• Retrospective analysis of 184,935 screening mammograms (DM and DBT) from 2010-2019.
• Identification and classification of 148 interval breast cancers (IBCs) by breast radiologists.
• Application of a deep-learning AI tool to assign risk scores to negative screening mammograms and evaluate its flagging and localization accuracy.

Main Results:

• Of 148 IBCs, 26% were minimal signs-actionable, 24% occult, and 22% minimal signs-non-actionable.
• AI most frequently flagged exams with missed-reading errors (90%) and minimal signs-actionable (89%).
• AI demonstrated higher accuracy in localizing mammographically visible IBCs (35-68%) compared to non-visible types (0-50%).

Conclusions:

• AI tools can effectively flag mammographically visible interval breast cancers (IBCs), including those with missed or minimal signs.
• AI accuracy in localization was significantly better for visible IBC types than for occult or true interval cancers.
• Findings suggest AI's potential to improve early detection of certain IBCs within annual screening programs.