Using deep learning models to triage screening mammograms can improve radiologist specificity without hurting sensitivity, according to new research published in Radiology. It also works as a way to help radiologists manage their heavy caseloads.
“Mammography is the only imaging modality shown to reduce breast cancer mortality in randomized trials,” wrote Adam Yala, department of electrical engineering and computer science at Massachusetts Institute of Technology in Cambridge, and colleagues. “Despite its benefits, challenges include variation in interpretive performance and the scarcity of specialized radiologists.”
Yala et al. explored data from more than 223,000 consecutive screening mammograms from January 2009 to December 2016 at a single academic medical center. After making exclusions as needed, data from more than 56,000 patients were used as a training set for the deep learning model. Data from more than 7,000 patients were used for both validating and testing the model.
“We emphasize that we split our data set by patients, and so each woman contributed mammograms to only one set,” the authors explained.
The deep learning system was then brought into a new simulated triage workflow that allowed radiologists to skip interpreting mammograms identified as being “cancer free.” Overall, the simulated workflow resulted in radiologists reading 80.7% of all screening mammograms, obtaining a sensitivity of 90.1% and specificity of 94.2%. When reading all mammograms, on the other hand, the radiologists had a sensitivity of 90.6% and specificity of 93.5%.
“We developed a deep learning model to triage mammograms as cancer free and showed that our model could improve radiologist efficiency and specificity without harming sensitivity,” the authors wrote. “This work is a first step to using deep learning to triage mammograms in routine clinical care.”
Yala and colleagues also noted that their deep learning model was effective for women with a range of breast densities, a clear sign that it would be “widely applicable to diverse patient populations.”