Justin D. Krogue MD, Kaiyang Cheng, Kevin Hwang MD, Eugene Ozhinsky PhD, Sharmila Majumdar PhD, Valentina Pedoia PhD
Overview: Hip fractures represent an important cause of morbidity, with more than 250,000 annually in the U.S. alone at a cost of $10-15 billion. Accurate and timely diagnosis is critical, as outcomes deteriorate with increased time to mobilization and management differs by fracture type. This research aims to automatically identify hip fractures from hip radiographs using neural networks in order to prioritize radiologist’s reading queues and to provide early notification of important injuries to the emergency room and orthopedic physicians.
Multi-domain data comprising imaging (x-ray, CT and MRI), radiologists’ reports and surgical history is utilized to provide the ground truth. We show that semi-automated pipelines using deep learning can identify hip fractures with performance comparable to radiologists (fracture identification sensitivity 92.0%, specificity 94.0%, multi-subclass accuracy 88.6%). Fully-automatic annotation pipelines are being developed to localize hip from the radiograph and using Natural Language Processing (NLP) models to read radiology reports.
The models have the potential to quickly filter radiographs to classify subjects at a high risk of Hip Fracture as well as identify relevant features that may play a role in the development and presence of Hip Fracture through a heat-map analysis.
The authors would like to acknowledge the contributions of Thomas Vail to this project.
