As deep learning makes its way into clinical practice, implementing it in diagnostic imaging may produce results on par with radiologists. A recent study published in Radiology demonstrated that deep-learning bone age assessment models analyzing hand radiographs produced results as accurate as a radiologist.
The researchers, led by David Larson, MD, MBA, from the Stanford University School of Medicine, found a deep learning convolutional network model could estimate bone maturity as well as human counterparts.
"We emphasize that automated assessment of bone age is likely among the easiest applications for deep learning in medical imaging; more complex applications may be less accurate, less clinically relevant or less feasible to develop," Larson and colleagues wrote.
The researchers obtained 14,036 left-hand radiographs and reports from two children's hospitals: Lucile Packard Children's Hospital at Stanford University and Children's Hospital Colorado. They then split the radiographs into two sets. The first set of 200 examinations (50 percent male, 50 percent female) used a mean of bone age estimates from the clinical report and three additional human reviewers as the reference standard. The second set, on the other hand, contained a total of 913 deidentified images obtained from the Digital Hand Atlas, developed by the University of Southern California Image Processing and Informatics Laboratory.
Performance of the deep learning model was determined comparing the root mean square and mean absolute difference between the model estimates and the reference standard bone ages, according to the researchers.
"Ninety-five percent limits of agreement were calculated in a pairwise fashion for all reviewers and the model," the researchers wrote. "The RMS of a second test set composed of 913 examinations from the publicly available Digital Hand Atlas was compared with published reports of an existing automated model."
The mean difference between bone age estimates of the model and reviewers was zero years, according to study results. The estimates of the model, clinical report and three reviewers were within 95 percent limits of agreement, according to study results. In accordance with these findings, RMS for the Digital Hand Atlas data set was 0.73 years compared with 0.61 of previous reported models.
Additionally, clinical classification of bone age as advanced, normal or delayed wasn't found to be very different for any of the reviewers or the model. However, researchers did find that the deep learning model increased in accuracy as the data sets increased.
"On the basis of these assessments, we conclude that an automated model for assessment of bone age based on a convolutional neural network can have an accuracy similar to that of current state-of-the-art automated models by using feature-extraction techniques," the researchers wrote. "Our results suggest potential broad applicability of deep-learning models for a variety of diagnostic imaging tasks without requiring specialized subject matter knowledge or image-specific software engineering."
A recent graduate from Dominican University (IL) with a bachelor’s in journalism, Melissa joined TriMed’s Chicago team in 2017 covering all aspects of health imaging. She’s a fan of singing and playing guitar, elephants, a good cup of tea, and her golden retriever Cooper.