A deep learning algorithm developed for processing digital screening mammograms, Mirai, was able to detect interval breast cancers and identify women who would benefit from more frequent screenings, according to the results of a UK retrospective study published in Radiology.
“If we called back 20% of women for supplemental imaging, we’d have to find the capacity to offer contrast-enhanced mammography or magnetic resonance imaging (MRI) to 440,000 women,” explained co-author Fiona J. Gilbert, MBChB, Professor of Radiology, Department of Radiology, University of Cambridge, United Kingdom, and Honorary Consultant Radiologist at Addenbrooke's Hospital. “Identifying women at an increased risk of developing breast cancer is a complex, multifactorial problem. The goal is to accurately identify the women most likely to have an interval cancer while minimizing the volume of supplemental imaging performed.”
Background and Study Methods
Interval breast cancers tend to be larger or more aggressive than other breast cancers as they develop quickly between screenings. Artificial intelligence (AI) has already demonstrated benefit in assisting radiologists in identifying breast cancers.
The authors of the new study sought to assess the performance of an AI-based model for personalizing breast cancer screenings in a triennial screening program (the UK breast cancer screening program performs mammograms for women aged 50 to 70 years old every 3 years).
Mirai processed the mammograms of 134,217 women that were collected between January 2014 and December 2016 across two sites and two primary mammography systems. Interval cancers were compared against a reference standard of no cancer diagnosis within 40 months of negative screening. A generalized risk score was issued for the chances of the woman developing an interval breast cancer, and women with higher scores were identified as possibly benefiting from supplemental screening of MRI or contrast-enhanced mammography, or shortened screening intervals.
Study Findings
A total of 524 interval cancers were identified by Mirai. No performance differences were noted between the 1-, 2-, and 3-year interval cancer predictions, age quartiles, or breast densities.
The area under the curve was 0.72 (95% confidence interval [CI] = 0.62–0.71) for the 1-year interval cancer prediction, 0.67 (95% CI = 0.64–0.70) for 2 years, and 0.67 (95% CI = 0.65–0.70) for 3 years, showing that the model performed better at predicting interval cancers within 12 months of screening than at later time points.
By age quartiles, the C-index was 0.67 (95% CI = 0.62–0.71) for women younger than age 52; women aged 52 to 58 years had a C-index of 0.70 (95% CI = 0.65–0.75); women aged 59 to 65 had a C-index of 0.71 (95% CI = 0.67–0.75); and women 66 or older had a C-index of 0.71 (95% CI = 0.67–0.75).
By BI-RADS categories, those in category A had a C-index of 0.70 (95% CI = 0.62–0.78), 0.69 for category B (95% CI = 0.65–0.73), 0.68 for category C (95% CI = 0.64–0.71), and 0.67 for category D (95% CI = 0.62–0.73), indicating that the model performed better in women with less dense breast tissue.
Three-year risk scores predicted 3.6% of interval breast cancers in women with the highest 1% of scores, 14.5% of cancers in the highest 5%, 26.1% of cancers in the highest 10%, and 42.4% of cancers in the top 20% of scores.
“Our results suggest that further workup of mammograms within the top 20% of scores could yield 42.4% of interval cancers, meaning that Mirai could be used to identify women for supplemental imaging or a shortened screening interval, instead of or in addition to breast density,” said lead author Joshua W.D. Rothwell, MBBS/PhD Student at the University of Cambridge.
Disclosure: For full disclosures of the study authors, visit pubs.rsna.org.
