Browsing by Author "Girod, Miguel M. "

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    A Deep Learning Algorithm to Detect Proximal Humerus Fractures on Radiographs
    (2025) Sperling, John William; Yang, Linjun; Girod, Miguel M. ; Saniei, Sami; Kaji, Elizabeth S. ; Grove, Austin F. ; Khela, Monty; De Marinis Acle, Rodrigo Ignacio; Sanchez Sotelo, Joaquin
    Background Proximal humerus fractures are one of the most common fractures in the elderly. Management of these injuries varies depending on the fracture pattern. It has become recognized that agreement is poor when these fractures are classified according to most traditional schemes. Deep learning (DL) offers the promise to improve recognition of specific fracture patterns. A first, necessary step in the development of DL pipelines to automatically classify fractures into specific patterns is to automatically detect that the proximal humerus is actually fractured, so that a fracture pattern DL classifier is not applied to non-fracture x-rays (normal, arthritic, and other). The purpose of this study was to develop a reliable and trustworthy DL approach to detect proximal humerus fractures on radiographs.MethodsAfter obtaining a patient cohort and reviewing and labeling their associated shoulder images, radiographs of fractured (n=996) and non-fractured (n=607) proximal humerus were used for this study. For model training, a random search was performed to fine-tune and determine the training hyper-parameters. All radiographs were split into six sets. The first five sets were used for model development using five-fold cross-validation, and the sixth set was used for internal model testing. The best-performing model was determined using the F1 score on the sixth set and was further validated using an external test set of 116 separate additional radiographs obtained after proximal humerus fractures. Saliency maps were generated for visual understanding of the DL model.ResultsThe best-performing model from the five-fold cross-validation achieved an accuracy of 0.972 and an F1 score of 0.969 on the hold-out test set. When further validated using the external test set of 116 radiographs of fractured proximal humeri, the model achieved an accuracy/sensitivity of 0.966, misclassifying only 4 of the 116 fractures. The saliency maps showed that the model focused on the perimeter of the humeral head, including the greater tuberosity and the surgical neck, when detecting the fracture.ConclusionThe DL algorithm developed in this study displayed robust and trustworthy performance when detecting the presence of a proximal humerus fracture on radiographs. This algorithm could be implemented in busy emergency department practices so that proximal humerus radiographs with a high probability of fracture can be flagged for dedicated review. Additionally, this algorithm provides the first step to further develop other AI tools to better understand and manage proximal humerus fractures.