A deep-learning model classified acute and nonacute pediatric elbow abnormalities on radiographs in trauma with 88 percent accuracy, according to new research published in Radiology: Artificial Intelligence.
“In high-volume emergency departments or urgent care centers without trained on-site pediatric radiologists, there is a strong need for quick and correct triage of the patient for either orthopedic evaluation or discharge,” wrote lead by Jesse C. Rayan, MD, of the Massachusetts General Hospital, and colleagues. “Binomial identification of elbow abnormality has the potential to simplify point-of-care triage in settings lacking immediate access to a trained pediatric radiologist.”
Rayan and colleagues sought to determine the feasibility of using deep learning with a multiview approach for pediatric elbow abnormalities on radiographs—similar to how radiologists review multiple images at their workstations.
“To our knowledge, no studies to date have experimented with [convolutional neural networks] (CNN) application in pediatric elbow examinations and tested the ability for differentiating abnormalities from normal growth centers,” Rayan and colleagues wrote.
Rayan et al. studied a total of 21,456 radiographic studies that contained more than 58,000 images of the elbow over a four-year period, between January 2014 through December 2017.
The studies were binomially classified as either positive or negative for acute or subacute traumatic abnormality and then randomly divided them into a training set containing 20,350 studies and a validation set containing 1,106 studies. The researchers combined a convolution neural network and recurrent neural network to interpret an entire series of three radiographs together.
“Fractures were successfully identified with deep learning in skeletally immature patients with open growth plates, and these open growth plates were distinguished from true abnormality,” the researchers wrote.
The researchers found their deep-learning model was able to detect abnormalities with 88 percent accuracy, with a sensitivity of 91 percent and specificity of 84 percent. AI missed the following number of abnormalities:
- 1 supracondylar fracture out of 241 cases.
- 1 lateral condylar fracture out of 88 cases.
- 15 elbow effusions without fracture out of 77 cases.
- 37 other abnormalities out of 184 cases.
The researchers noted their AI can effectively binomially classify acute and nonacute findings on pediatric elbow radiography in trauma. Furthermore, their method was unique, as they applied a recurrent neural network to classify an entire radiographic series, rather than a single radiographic image.
The model can arrive at a decision based on all views, similar to that of a human radiologist, and will continue to improve with further modifications.