Deep learning improves DBT's efficiency

Sunday, December 1 | 10:45 a.m.-10:55 a.m. | SSA01-01 | Room S406A

In this presentation, researchers will describe how using a deep-learning algorithm with digital breast tomosynthesis (DBT) can reduce radiologists' interpretation times.

Presenter Flora Gilboa-Solomon of IBM Research in Haifa, Israel, will share findings from a study that demonstrated that artificial intelligence (AI) can filter normal DBT exams from a radiologist's worklist without increasing the false-negative rate.

The study included 5,000 women presenting for screening DBT between 2013 and 2017; 12,500 DBT exams were performed. The team trained a deep-learning network to distinguish between benign and malignant breast lesion features with the goal of using it to filter out normal exams and maintain a false-negative rate equivalent to that of radiologists.

Gilboa-Solomon's team found that the network successfully filtered 37% of normal exams with 97% specificity, with an area under the receiver operating curve of 0.84 for the task.

"One important practical issue related to DBT implementation is the longer interpretation time," the group concluded. "Reducing the workload of reading normal exams can improve radiologist's efficiency."