AI helps pathologists diagnose difficult breast cancer cases
A new machine learning system created by UCLA researchers may help doctors classify breast cancers that are notoriously difficult to diagnose, according to an Aug. 9 study published in JAMA Network Open.
Preinvasive lesions, including some types of atypia and ductal carcinoma in situ (DCIS) can indicate higher cancer risk, wrote corresponding author Joann G. Elmore, MD, MPH, with UCLA’s Division of General Internal Medicine and Health Services Research, but diagnostic disagreements are “remarkably high” for such lesions. Machine learning has shown great promise for differentiating breast cancers, but studies have largely focused solely on tumor detection.
“Medical images of breast biopsies contain a great deal of complex data and interpreting them can be very subjective,” Elmore added, in a prepared statement. “Distinguishing breast atypia from ductal carcinoma in situ is important clinically but very challenging for pathologists. Sometimes, doctors do not even agree with their previous diagnosis when they are shown the same case a year later.”
Elmore and colleagues created their platform to automatically read tissue distribution features and structure features, training the computer model using 240 breast biopsies from multiple Breast Cancer Surveillance Consortium registries. They selected samples that varied by breast density, diagnosis, patient age and biopsy type; 3 expert pathologists reviewed and categorized the biopsies as benign, atypia, DCIS or invasive cancer.
In order to test their system, readings were compared to independent diagnoses made by 87 practicing pathologists.
“Our methods differ from prior work in that we attempted to emulate the behavior of pathologists as they interpret these cases by tackling successive binary decisions that were sequentially more challenging on the diagnostic difficulty scale,” Elmore et al. noted.
Overall, the AI program fell just short of trained pathologists for differentiating cancer from non-cancer: machine learning distribution features scored a 0.94, structure features a 0.91 and pathologists 0.98. However, the model achieved its primary goal of differentiating DCIS from atypia. It registered a sensitivity between 0.88 and 0.89 compared to the doctors’ average sensitivity of 0.70.
According to the release, classifying DCIS from atypia is “the greatest challenge in breast cancer diagnosis.”
“These results are very encouraging,” Elmore said. “There is low accuracy among practicing pathologists in the U.S. when it comes to the diagnosis of atypia and ductal carcinoma in situ, and the computer-based automated approach shows great promise.”
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.