An AI algorithm for breast cancer screening has potential to enhance the performance of digital breast tomosynthesis (DBT), reducing interval cancers by up to one-third, according to a study published today in Radiology.
Interval breast cancers tend to have poorer outcomes due to their more aggressive biology and rapid growth. DBT, or 3D mammography, can improve visualization of breast lesions and reveal cancers that may be obscured by dense tissue. Because DBT is relatively new as an advanced screening technology, long-term data on patient outcomes are limited in institutions that have not transitioned to DBT until recently.
“Given the lack of long-term data on breast cancer-related mortality measured over 10 or more years following the initiation of DBT screening, the interval cancer rate was often used as a surrogate marker,” explained study author Manisha Bahl, MD, MPH, breast imaging division quality director and co-service chief at Massachusetts General Hospital and associate professor at Harvard Medical School. “Lowering this rate is assumed to reduce breast cancer-related morbidity and mortality.”
In a study of 1,376 cases, Dr. Bahl and her colleagues retrospectively analyzed 224 interval cancers in 224 women who had undergone DBT screening. On those DBT exams, the AI algorithm (Lunit INSIGHT DBT v1.1.0.0) correctly localized 32.6% (73/224) of cancers that were previously undetected.
“My team and I were surprised to find that nearly one-third of interval cancers were detected and correctly localized by the AI algorithm on screening mammograms that had been interpreted as negative by radiologists, highlighting AI’s potential as a valuable second reader,” Dr. Bahl said.
According to the researchers, the Radiology study may represent the first published research to specifically examine AI assistance in detecting interval cancers on screening DBT exams.
“Several studies have explored the use of AI to detect interval cancers on screening two-dimensional digital mammography exams, but to our knowledge no previously published literature has focused on the use of AI to detect interval cancers on DBT,” Dr. Bahl explained.
Improved Lesion-Level Accuracy
To avoid overestimating the sensitivity of the AI algorithm, Dr. Bahl’s team employed a lesion-specific analysis that “credits” the AI algorithm only when it correctly identifies and localizes the exact site of the cancer.
“In contrast, an exam-level analysis gives AI credit for any positive exam, even if its annotation is incorrect or unrelated to the actual cancer site, which may inflate the algorithm’s sensitivity,” Dr. Bahl said. “Focusing on lesion-level accuracy provides a more accurate reflection of the AI algorithm’s clinical performance.”