Deep Learning Approach for Automatic Wrist Fracture Detection Using Ultrasound Bone Probability Maps

Wrist fractures are currently examined using radiograms, resulting in undesirable radiation exposure in children. Ultrasound is fast, safe, and highly sensitive to fractures, making it ideally suited for wrist examination in emergency departments (ED). However, ultrasound images are difficult to interpret, resulting in high variability in assessment depending on the reader’s expertise. We developed a new machine learning (ML) technique to detect fractures from 3D ultrasound (3DUS). We generate bone probability maps using local phase (LP) information in each ultrasound frame, combine these into a feature sequence, and analyze the same to predict the probability of fracture using three variants of recurrent neural networks (RNN): vanilla RNN, long short-term memory (LSTM), and gated recurrent unit (GRU) models. This approach was validated on 30 3DUS volumes, each of which was assessed by a radiologist for the presence of a fracture. RNN, LSTM and GRU gave 83%, 90%, and 87% accuracy when compared to clinical assessment by expert musculoskeletal radiologist, with GRU giving the most balanced sensitivity and specificity. The automatic assessment technique is reliable in detecting wrist fractures from 3D ultrasound and can be used as a valuable ED triage tool for fracture detection.