Fine-Tuned Deep Convolutional Networks for the Detection of Femoral Neck Fractures on Pelvic Radiographs: A Multicenter Dataset Validation

In this study, we aim to provide a deep convolutional network based femoral neck fracture detection system on radiographs for emergency patients. We retrospectively collected 1,491 frontal pelvic radiographs from three institutions and assigned them to the following data sets: primary dataset (710 radiographs, to fine-tune and validate the initial model called the Digital Radiography Fracture Detection System [DR-FDS]), internal test set 1 (189 radiographs) and 2 (235 radiographs), and external test set 1 (189 radiographs) and 2 (168 radiographs). Per-bounding box recall and precision and per-image sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC) were computed. We randomly extracted 300 radiographs from the above test sets and compared their effect on the diagnostic accuracy and efficiency of fine-tuned model-assisted and unassisted clinicians. The fine-tuned DR-FDS showed a better overall performance in detecting femoral neck fractures than did the initial DR-FDS. The fine-tuned DR-FDS achieved AUC values of 0.9526 (95%CI, 0.9048-0.9767) and 0.9633(95%CI, 0.9346-0.9797) in internal test sets 1 and 2. In external test sets 1 and 2, this model also achieved promising results with AUC values of 0.9231 (95%CI, 0.8779-0.9520), and 0.9937 (95%CI 0.9739-0.9985), respectively. The clinicians showed a statistically significant increase in specificity, sensitivity, and accuracy for the identification of minimal/undisplaced fracture and a decrease in the average reading time. The object detection model that is fine-tuned has high sensitivity and specificity and the universal ability to detect and locate femoral neck fractures on pelvic radiographs.