The study, published in the Journal of the National Cancer Institute, found that AI was able to identify "mammographically-visible" types of interval cancers earlier by flagging them at the time of screening. These include tumors that are visible on mammograms but not detected by radiologists, or have very subtle signs on mammography that are easy to miss because the signs were faint or arguably below the level of detection by the human eye.
Researchers estimate that incorporating AI into screening could help reduce the number of interval breast cancers by 30%.
"This finding is important because these interval cancer types could be caught earlier when the cancer is easier to treat," said Dr. Tiffany Yu, assistant professor of Radiology at the David Geffen School of Medicine at UCLA and first author of the study. "For patients, catching cancer early can make all the difference. It can lead to less aggressive treatment and improve the chances of a better outcome."
While similar research has been conducted in Europe, this study is among the first to explore the use of AI to detect interval breast cancers in the United States. Researchers point out that there are key differences between the U.S. and European screening practices. In the U.S., most mammograms are performed using digital breast tomosynthesis (DBT), often called 3D mammography, and patients are typically screened every year. In contrast, European programs usually use digital mammography (DM), often called 2D mammography, and screen patients every two to three years.
The retrospective study analyzed data from nearly 185,000 past mammograms from 2010–2019 that included both DM and DBT. From the data, the team looked at 148 cases where a woman was diagnosed with interval breast cancer.
Radiologists then reviewed these cases to determine why the cancer wasn’t spotted earlier. The new study adapted a European classification system to categorize the interval cancers. They include: Missed reading error, minimal signs–actionable, minimal signs–non–actionable, true interval cancer, occult (which is truly invisible on mammogram), and missed due to a technical error.
Researchers then applied a commercially available AI software called Transpara to the initial screening mammograms performed before the cancer diagnosis to determine if it could detect subtle signs of cancer that were missed by radiologists during initial screenings, or at least flag them as suspicious. The tool scored each mammogram from 1 to 10 for cancer risk. A score of 8 or higher was considered flagged as potentially concerning.
Key Findings:
- The team found that the AI flagged 76% of the mammograms that had been originally read as normal but were later linked to an interval breast cancer.
- It flagged 90% of missed reading error cases where the cancer had been visible on the mammogram but missed or misinterpreted by the radiologist.
- It caught about 89% of minimal-signs-actionable cancers that showed very subtle signs and could reasonably have been acted upon, as well as 72% of those with minimal-signs-non-actionable that were likely too subtle to prompt action.
- For cancers that were occult or completely invisible on the mammogram, the AI flagged 69% of cases.
- It was somewhat less effective at identifying true interval cancers, those that were not present at the time of screening but developed later, flagging about 50% of those.
"While we had some exciting results, we also uncovered a lot of AI inaccuracy and issues that need to be further explored in real-world settings," said Dr. Hannah Milch, assistant professor of Radiology at the David Geffen School of Medicine and senior author of the study. "For example, despite being invisible on mammography, the AI tool still flagged 69% of the screening mammograms that had occult cancers. However, when we looked at the specific areas on the images that the AI marked as suspicious, the AI did not do as good of a job and only marked the actual cancer 22% of the time."
Larger prospective studies are needed to understand how radiologists would use AI in practice and address key questions, such as how to handle cases where AI flags areas as suspicious that aren’t visible to the human eye, especially when the AI isn’t always accurate in pinpointing the exact location of cancer.
"While AI isn't perfect and shouldn't be used on its own, these findings support the idea that AI could help shift interval breast cancers toward mostly true interval cancers," Yu added. "It shows potential to serve as a valuable second set of eyes, especially for the types of cancers that are the hardest to catch early. This is about giving radiologists better tools and giving patients the best chance at catching cancer early, which could lead to more lives saved."
The work was supported in part by the National Institutes of Health, the National Cancer Institute, the Agency for Healthcare Research and Quality and Early Diagnostics Inc.
Yu TT, Hoyt AC, Joines MM, Fischer CP, Yaghmai N, Chalfant JS, Chow L, Mortazavi S, Sears CD, Sayre J, Elmore JG, Hsu W, Milch HS.
Mammographic classification of interval breast cancers and artificial intelligence performance.
J Natl Cancer Inst. 2025 Apr 18:djaf103. doi: 10.1093/jnci/djaf103