AI algorithm flags interval breast cancer cases

AI could help tackle the challenge of identifying interval breast cancers, according to research presented November 30 at the RSNA 2023 annual meeting.

In her talk, Tiffany Yu, MD, from the University of California, Los Angeles presented findings demonstrating how an AI algorithm correctly flagged minimal signs on digital mammography and digital breast tomosynthesis (DBT), as well as reading error misses.

“AI can potentially play a role in identifying and reducing the interval breast cancer rate in U.S.-based screening programs,” Yu said.

While recent studies based in Europe highlight how AI can reduce interval breast cancers, Yu noted that there are no comparable U.S. studies in this area to her or her colleagues’ knowledge. She added that standardization of classifying interval breast cancer is difficult due to variable definitions and reporting. This includes screening intervals, technology used, and the number of readers per exam.

She and her team reviewed and classified interval breast cancers to evaluate a commercially available AI model’s (Transpara 1.7.1) performance within U.S. screening guidelines. They acquired data from full-field digital mammography and DBT exams between 2010 and 2019. They also excluded exams with BI-RADS ratings of 0, 3, 4, 5, and 6.

For the study, eight fellowship-trained breast radiologists retrospectively classified the interval cancers as one of the following: true interval cancer, minimal signs-non-actionable, minimal signs-actionable, technical miss, reading error miss, or occult. The radiologists had varying degrees of experience ranging from three to 24 years.

In total, the team included 150 interval cancers in 150 women in its study from a dataset of 184,935 screening mammograms (65% digital mammography, 35% DBT exams).

It found that AI flagged 63.3% of the overall interval cancer cases and had varying performance in flagging different classifications. However, it achieved the best performance on minimal signs-actionable and missed-reading error cases.

Performance of AI in flagging breast imaging findings in interval cancer cases
Interval cancer classification Number of cancers classified AI flag proportion
Missed reading-error 25 94%
Minimal signs-actionable 38 89%
Minimal signs-nonactionable 35 71%
Occult 36 68%
True interval cancers 9 44%
Missed-technical error 7 40%

The researchers also found that women aged 40-49 years were disproportionately affected by occult interval cancers at 42% compared to women aged 50 years and older with minimal signs-actionable (27%, p = 0.007). Additionally, they found significant ties between interval cancers and breast density, with 45% of cases being associated with density category C as set by the American College of Radiology.

The team also reported that triple-negative cancers comprised 13% of the study sample, with the greatest proportion being true interval cancers (25%, p = 0.003). Finally, it found that interval cancers were localized, with 86% being tumor, node, metastasis (TNM) stage 0 to 2B.

Yu said that future research will explore how the AI algorithm performs with different lesion types, including architectural distortion cases and calcifications.

Page 1 of 568
Next Page