Deep learning can improve the accuracy and efficiency of digital breast tomosynthesis (DBT) examinations, according to new findings published in Radiology: Artificial Intelligence.
The study’s authors developed and trained a deep learning system to identify suspicious findings on DBT images, hoping it could help reduce the longer reading times associated with screening patients with DBT. A group of 24 radiologists then read 260 DBT examinations with and without assistance from the deep learning system.
The move made an immediate impact, increasing sensitivity from 77% to 85% and specificity from 62.7% to 69.6%. In addition, the recall rate for noncancers decreased from 38% to 30.9% and reading times decreased from 64.1 seconds to 30.4 seconds. The area under the ROC curve increased from 0.795 to 0.852.
“Overall, readers were able to increase their sensitivity by 8 percent, lower their recall rate by 7 percent and cut their reading time in half when using AI concurrently while reading DBT cases compared to reading without using AI,” lead author Emily F. Conant, MD, professor and chief of breast imaging from the department of radiology at the Perelman School of Medicine at the University of Pennsylvania in Philadelphia, said in a prepared statement.
Conant added that these findings could help get reading times “back to about the time it takes to read digital mammography-alone exams.”
Conant and colleagues explained in the study that using a deep learning system such as the one they developed could represent a challenge initially for radiologists, but it is well worth the effort.
“The concurrent use of AI from the start of DBT reading will be a change and will possibly require a learning curve for most breast imagers,” the authors wrote. “However, because concurrent use of AI yields significant improvements in both efficiency and accuracy, such a reading protocol with AI should be considered by all breast imagers.”
The team also added that the study had certain limitations. Radiologists “may behave differently” in reader studies than they do on a typical work day, for instance, and more details about the patient outcomes could help researchers “fully understand the clinical impact of this AI system.”
“These limitations and others might be addressed with a larger prospective study,” the authors wrote.