Deep learning tools can help emergency room physicians and radiologists, even those without specific musculoskeletal training, make faster and more accurate diagnoses of fractures, according to a new study presented Wednesday.
“Identifying bone fractures on radiographs is one of the most common tasks that medical trainees, emergency physicians, and radiologists have to do in the emergency department,” said George Fu, MD, resident in the Department of Radiology at Case Western Reserve University School of Medicine. “However, due to the large volume of exams performed and a relatively fewer number of musculoskeletal subspecialty trained radiologists, these exams are usually preliminarily read by radiology trainees and emergency physicians, and signed off by a non-musculoskeletal trained radiologist.”
According to Dr. Fu, that difference can lead to accuracy issues and delays in patient care. The study looked at whether a deep learning-based AI algorithm, which was developed for the purpose of aiding physicians in detecting fractures, could be used to improve performance.
Using Technology Can Improve Accuracy And Shorten Reading Time
The research team used a deep learning tool that was previously developed to detect fractures using 132,000 appendicular skeleton radiographs from multiple institutions across nine countries.
For this work, Dr. Fu and his team used the tool to evaluate 2,626 radiographs from four centers at one institution in Ohio through a multi-reader retrospective study where 24 readers (eight each of emergency physicians, non-musculoskeletal radiologists, and musculoskeletal radiologists) were asked to identify fractures with and without the deep learning tool aid.
The stand-alone accuracy of the deep learning model was more than 98%, with sensitivity and specificity near perfect. Average reading time was shortened by 7.1 seconds or 27% per exam with the help of the deep learning tool.
“Deep learning-based tools have the potential to be used to improve physician performance in fracture detection on radiographs and therefore may help improve patient care,” Dr. Fu said. “As a radiologist and also someone who is passionate about computer technology, I hope to be able to use technological solutions to improve the care we provide for patients, making sure we don’t miss important findings and getting diagnoses to clinicians faster.”
Additional Research Needed Before Clinical Implementation
According to Dr. Fu, improving diagnostic accuracy reduces the number of fractures that are missed on radiographs. By shortening the amount of time it takes to read these studies, AI-based tools have potential to relieve the workload burden on radiologists.
He noted that there needs to be more study of the application of AI based tools in real clinical settings to assess how they improve radiology workflows, radiologist and trainee performance, case list triaging, and report turnaround time.
“We also hope our study can serve as a guide for other institutions and radiology groups to conduct studies to evaluate AI tools that they employ or hope to employ in their practice,” he said.
Access the presentation, “Assessing the Potential of a Deep Learning Tool To Improve Fracture Detection by Radiologists and Emergency Physicians on Extremity Radiographs,” (W6-SSER02-1) on demand at Meeting.RSNA.org.