Researchers from NYU Langone Health trained and assessed an AI system utilizing an ensemble of models based on both full-field digital mammography (FFDM) and DBT images. The algorithm was trained and tuned on a dataset of nearly 500,000 FFDM and/or DBT images.
An ensemble model that combined predictions from DBT, synthetic mammography, and FFDM produced an area under the curve of 0.917 in a test set of 19,684 exams for identifying cases with malignant lesions. And when triaging 30%, 40%, or 50% of women with the lowest AI scores into a workflow that didn’t require radiologist interpretation, the algorithm would have yielded a false-negative rate of 0%, 0%, and 0.016%, respectively, according to presenter Dr. Laura Heacock and colleagues.
They concluded that up to 40% of screening DBT studies could be removed from the radiologist workflow with a 0% false-negative rate and 29% lower recall rate.
“An AI system as a standalone DBT interpreter could eliminate low probability of malignancy cases from the radiologist worklist, resulting in fewer false positives/recalls,” the authors wrote in their abstract.
Get all of the details by sitting in on this Thursday morning session.
