Ownership Interest: Stock in Imagen Technologies (Daluiski, Hotchkiss, Hanel, Chopra, Lindsey) Salary: Imagen Technologies (Chopra, Lindsey) Deep convolutional neural networks (CNNs) can accurately detect and localize distal radius fractures in radiographs better than inexperienced clinicians. Deep CNNs have had tremendous success in detecting, localizing, and identifying objects in many computer vision domains in which large labeled datasets are available for model training. We developed a novel CNN model and trained it on a dataset of 36,408 wrist radiographs labeled by two experienced hand surgeons (AD, RH) for the presence and location of fractures of the distal radius. The model treats the problem as a semantic segmentation task which first localizes regions of interest in a radiograph and then uses a deep CNN to output a heatmap which represents pixelwise probabilities of a fracture. The labeled dataset was divided into training, development, and test sets consisting of 80%, 10%, and 10% of the images, respectively. An additional held-out set was prepared consisting of 2,215 consecutive wrist radiographs from a hospital collected between July and September, 2016. The “ground truth” labels of this dataset were determined by the majority opinion of three authors (AD, RH, DH). The ROC AUC (area under the curve) was calculated to report the diagnostic accuracy of the model. To further asses the impact of our model on clinicians who could potentially be aided by this technology, we conducted a controlled experiment using two radiology trainees, two physician extenders, and two practicing urgent care physicians. Each clinician was asked to indicate the presence or absence of distal radius fractures in 250 radiographs randomly sampled from the held-out set, and we compared their performance against the model. Of the 36,408 radiograph films 31% had fractures. Children less than 10 years of age, females over 50 years old, and left wrists were more likely to have a fracture. Our model had an AUC of 0.97 on the test set. On the held-out set our model had an AUC of 0.98, and a specificity of 93% at an operating point of 95% sensitivity, which is significantly more accurate than the tested trainees, physician extenders or practicing urgent care physicians (Figs. 32-1, 32-2). •Our proposed CNN-based semantic segmentation model produce heatmaps which accurately identify the presence and location of distal radius fractures.•The accuracy of the model was better than the tested trainees, physician extenders, and urgent care physicians, suggesting the proposed technology may be helpful in both residency training and urgent care clinical settings.Figure 32-2Example heatmaps produced by the model. The shading of each pixel indicates the model’s confidence that the pixel is within a distal radius fracture.View Large Image Figure ViewerDownload Hi-res image Download (PPT)