A new study – “Mammographic Classification of Interval Breast Cancers and Artificial Intelligence Performance” – conducted at the University of California, Los Angeles (UCLA) and published in in the Journal of the National Cancer Institute, investigates interval breast cancers (IBCs) – cancers diagnosed within 12 months after a negative screening mammogram – in a U.S.-based annual screening program.
The research aimed to classify IBCs based on their mammographic visibility and evaluate the performance of an artificial intelligence (AI) tool in detecting these cancers, particularly in a setting that uses both digital mammography (DM) and digital breast tomosynthesis (DBT).
The study analyzed 184,935 screening mammograms (65% DM, 35% DBT) from 49,244 women screened between 2010 and 2019 at a U.S. tertiary care academic center participating in the Athena Breast Health Network. From this cohort, 148 IBCs were identified in 148 women (mean age 61 ± 12 years), defined as ductal carcinoma in situ (DCIS) or invasive carcinomas diagnosed within 12 months of a negative mammogram (BI-RADS 1 or 2).
Eight fellowship-trained breast radiologists, with 3–24 years of experience, retrospectively classified these IBCs into six categories adapted from European guidelines: missed-reading error (visible but missed), minimal signs-actionable (subtle signs potentially detectable), minimal signs-non-actionable (subtle signs not reasonably detectable), true interval (not present at screening), occult (not visible on mammography but detectable by other modalities), and missed-technical error (missed due to technical issues like positioning).
A deep-learning AI tool (Transpara v1.7.1, ScreenPoint Medical) assigned risk scores (1–10) to 131 of the 148 negative index mammograms (17 were excluded due to breast implants or file conversion issues). Exams scoring ≥8 were considered “flagged” for elevated risk. Radiologists assessed whether AI markings accurately localized the IBC site, comparing performance across IBC types and correlating with patient and tumor characteristics.
The AI tool flagged 76% of the 131 scored mammograms, with the highest flagging rates for missed-reading errors (90%), minimal signs-actionable (89%), and minimal signs-non-actionable (72%). These mammographically visible types also received higher mean AI scores (9.5, 8.9, and 8.3, respectively) compared to non-visible types like true interval (7.3) and occult (7.9). AI accurately localized the cancer site in 47% of flagged exams, with the highest accuracy for missed-reading errors (68%) and minimal signs-actionable (62%), but lower for occult (22%) and missed-technical errors (0%). Overall, AI flagged 56% of mammographically visible IBCs versus 20% of non-visible ones, suggesting a potential 30% reduction in IBCs if AI-supported detection were implemented.
The study highlights the potential of AI to enhance the detection of mammographically visible IBCs, particularly those missed or with subtle signs, in a U.S. annual screening program using both DM and DBT. Unlike European studies, the lower rate of true interval cancers (6%) underscores the benefit of annual screening in reducing IBCs. However, AI’s ability to flag occult and true interval cancers, despite lower localization accuracy, raises questions about its predictive value and clinical utility, as radiologists may struggle to act on flagged exams without visible abnormalities.
Limitations include the retrospective design, small sample sizes for some IBC subtypes, and the use of a single AI vendor, which may limit generalizability. The mix of DM and DBT reflects real-world practice but complicates comparisons. Selection bias from the Athena study cohort and technical issues with unscored exams are additional constraints. Importantly, the study does not confirm whether AI-flagged exams would lead to earlier detection in practice, as radiologist recall decisions remain uncertain.
The study was supported by the National Institutes of Health, the Agency for Healthcare Research and Quality, and Early Diagnostics Inc., with no reported conflicts of interest. The data are available upon request from the corresponding author, Tiffany T. Yu, MD.