Object detection convolutional neural networks (CNNs) can detect and localize fractures in wrist x-rays, according to a new study published in Radiology: Artificial Intelligence.
“Interpretation errors on radiographs are contributed by human and environmental factors, such as clinician inexperience, fatigue, distractions, poor viewing conditions, and time pressures,” wrote Yee Liang Thian, department of diagnostic imaging at the National University of Singapore, and colleagues. “Automated analysis of radiographs by computers, which are consistent and indefatigable, would be invaluable to augment the work of emergency phsicians and radiologists.”
Object detection CNNs are different from the deep learning networks being used in other recent studies, the authors explained, because they can provide specific location information of the abnormality in question. A bounding box is drawn around the object in question, alerting specialists of where it can be found.
The researchers used more than 7,000 wrist x-rays for the study, with radiologists annotating all radius and ulna fractures with bounding boxes of their own. While 90 percent of the x-rays were used to train an object detection CNN, the other 10 percent were reserved for validation. The fracture localization models, designed for frontal and lateral images, were then tested on an unseen collection of more than 500 emergency department (ED) wrist x-rays.
Overall, the CNN detected and provided location information for more than 91 percent of all radius and ulna fractures on the frontal view and more than 96 percent on the lateral view. For the frontal view, the per-image sensitivity, specificity and AUC were 95.7 percent, 82.5 percent and 0.918, respectively. And for the lateral view, the per-image sensitivity, specificity and AUC were 96.7 percent, 86.4 percent and 0.933, respectively. Per-study sensitivity was 98.1 percent, per-study sensitivity was 72.9 percent and per-study AUC was 0.895.
“The ability to predict location information of abnormality with deep neural networks is an important step toward developing clinically useful artificial intelligence tools to augment radiologist reporting,” the authors wrote.
One limitation of these findings, according to the study, was that only radius and ulna fractures were included. Another limitation was that the researchers focused exclusively on ED x-rays.
“We excluded training and testing with orthopedic outpatient radiographs because of the large proportion of metallic implants in routine orthopedic outpatient radiographs of the wrist,” the team explained. “Including such radiographs may unintentionally teach the CNN to associate the presence of metallic implant with presence of a fracture, rather than discriminative features of the fracture per se.”
Michael has more than 16 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.