Closed limb fractures are highly prevalent in disaster rescue scenarios, frontline battlefields, and routine military training. Their diagnosis and treatment require significant expertise and timeliness, as missed or delayed diagnosis can lead to more severe secondary trauma. This study aimed to utilize deep learning and ultrasonography for identifying closed limb fractures, constructing an intelligent sonographic recognition model and evaluating its diagnostic value.

A total of 173 patients with X-ray confirmed closed limb fractures and 122 healthy volunteers were enrolled in. Collected ultrasound images were used for deep learning feature extraction. An intelligent diagnostic model for limb fractures based on DenseNet121 was constructed. Model performance was evaluated using accuracy, sensitivity, specificity, precision, and F1-Score.

For bone classification (6-class), the model achieved accuracy of 90%, sensitivity of 85%, and F1-Score 85%. For fracture/non-fracture classification (2-class), the results were 99% accuracy and 99% F1-Score.

The DenseNet121-based model demonstrated promising results in identifying closed limb fractures from ultrasound images. The model may hold potential for aiding rapid identification and triage in emergency settings.