Mammographic Classification of Interval Cancers Using Artificial Intelligence

Product: Transpara Company: ScreenPoint Medical

Assessing AI Performance in Mammographic Classification of Interval Cancers

Journal of the National Cancer Institute, 2025

Abstract

Background

European studies suggest that artificial intelligence (AI) can reduce interval breast cancers. Research on interval breast cancer classification and AI’s effectiveness in the United States, however, particularly using digital breast tomosynthesis and annual screening, is limited. We aimed to mammographically classify interval breast cancers and assess AI performance using a 12-month screening interval.

Methods

From digital mammography and digital breast tomosynthesis screening mammograms acquired between 2010 and 2019 at a US tertiary-care academic center, we identified interval breast cancers diagnosed less than 12 months after a negative mammogram. At least 3 breast radiologists retrospectively classified interval breast cancers as missed—reading error, minimal signs—actionable, minimal signs—nonactionable, true interval, occult, or missed—technical error. A deep-learning AI tool assigned risk scores ranging from 1 to 10 to the negative index screening mammograms, with scores of 8 or higher considered “flagged.” Statistical analysis evaluated associations among interval breast cancer types and AI exam scores, AI markings, and patient and tumor characteristics.

Results

From 184 935 screening mammograms (65% digital mammography, 35% digital breast tomosynthesis), we identified 148 interval breast cancers in 148 women (mean [SD] age = 61 [12] years). Of these, 26% were minimal signs—actionable, 24% were occult, 22% were minimal signs—nonactionable, 17% were missed—reading error, 6% were true interval, and 5% were missed—technical error (P < .001). AI scored 131 mammograms (17 errors excluded); it most frequently flagged exams with missed—reading error (90%), minimal signs—actionable (89%), and minimal signs—nonactionable (72%) (P = .02). AI localized mammographically visible types more accurately (35%-68%) than nonvisible types (0%-50%; P = .02).

Conclusion

AI more frequently flagged and accurately localized interval breast cancer types that were mammographically visible at screening (missed or minimal signs) compared with true interval or occult cancers.

Read full study