Detecting intertrochanteric hip fractures with orthopedist-level accuracy using a deep convolutional neural network

ObjectiveTo compare performances in diagnosing intertrochanteric hip fractures from proximal femoral radiographs between a convolutional neural network and orthopedic surgeons.Materials and methodsIn total, 1773 patients were enrolled in this study. Hip plain radiographs from these patients were cropped to display only proximal fractured and non-fractured femurs. Images showing pseudarthrosis after femoral neck fracture and those showing artificial objects were excluded. This yielded a total of 3346 hip images (1773 fractured and 1573 non-fractured hip images) that were used to compare performances between the convolutional neural network and five orthopedic surgeons.ResultsThe convolutional neural network and orthopedic surgeons had accuracies of 95.5% (95% CI=93.1-97.6) and 92.2% (95% CI=89.2-94.9), sensitivities of 93.9% (95% CI=90.1-97.1) and 88.3% (95% CI=83.3-92.8), and specificities of 97.4% (95% CI=94.5-99.4) and 96.8% (95% CI=95.1-98.4), respectively.ConclusionsThe performance of the convolutional neural network exceeded that of orthopedic surgeons in detecting intertrochanteric hip fractures from proximal femoral radiographs under limited conditions. The convolutional neural network has a significant potential to be a useful tool for screening for fractures on plain radiographs, especially in the emergency room, where orthopedic surgeons are not readily available.