Researchers used a convolutional neural network (CNN) to classify acute and nonacute findings in pediatric elbow x-rays, according to new research published in Radiology: Artificial Intelligence. The team’s recurrent neural network was able to interpret an entire series of images together, mimicking the decision-making process of a human radiologist.
“Use of CNNs in the triage of radiologic studies has been suggested, but few studies have in fact studied their feasibility for such a task,” wrote Jesse C. Rayan of the E.B. Singleton Department of Pediatric Radiology at Texas Children’s Hospital in Houston, Texas, and colleagues. “To our knowledge, no studies to date have experimented with CNN application in pediatric elbow examinations and tested the ability for differentiating abnormalities from normal growth centers.”
The authors reviewed data from more than 21,000 pediatric elbow examinations performed from January 2014 to December 2017. Though demographic data was not available for studies from 2014, the mean age of the patients from the other three years was 7.2 years old. Fifty-seven percent of the patients from those three years were male. The team’s training set was made up of more than 20,000 studies, and the remaining studies were used as a validation set.
Overall, the CNN’s area under the receiver operating characteristic curve was 0.95. Accuracy was 88 percent, sensitivity was 91 percent and specificity was 84 percent.
“Our data demonstrate that deep learning can effectively binomially classify acute and nonacute findings on pediatric elbow radiographs in the setting of trauma,” the authors wrote.
Rayan and colleagues also noted that they applied a recurrent neural network that could classify an entire series of images instead of just making a classification based on a single image.
“This is analogous to how radiologists take into account all views before arriving at a diagnosis,” they wrote.