Development of an Artificial Intelligence System for Hip Fracture Detection: A YOLOv8 Model Performance Study for Junior Orthopedic Surgeons

Purpose: Hip fractures represent a critical orthopedic emergency in the geriatric population; diagnostic delays or inaccuracies may result in severe morbidity and mortality. Contemporary artificial intelligence technologies demonstrate potential for precise and rapid radiographic interpretation, particularly in resource-constrained healthcare environments with limited availability of specialists. We aimed to develop and validate the diagnostic performance of a YOLOv8-based deep learning model by junior orthopedic surgeons for the detection of hip fractures, categorizing images into three classifications: normal anatomy, femoral neck fractures, and intertrochanteric fractures. Methods: This retrospective study analyzed 2,035 anteroposterior hip radiographs from 942 patients. The YOLOv8 architecture was implemented using Google Colab with standardized hyperparameters. The dataset was stratified into training, validation, and testing sets. The performance evaluation utilized mean average precision ([email protected]), F1 score, precision, recall, sensitivity, specificity, and confusion matrix analysis. Results: The YOLOv8 model achieved an [email protected] of 0.879 and a maximum F1 score of 0.86. The model demonstrated a maximum precision, confidence threshold, and maximum recall of 1.00, 0.961, and 0.91, respectively, at a confidence threshold of 0.000. The sensitivity values were 97.7%, 87.0%, and 95.9% for intertrochanteric fractures, femoral neck fractures, and normal anatomy, respectively. The specificity ranged from 97.1% to 99.0% across all classifications, indicating exceptional screening accuracy, particularly for normal anatomy and intertrochanteric fractures. Conclusions: The YOLOv8 model demonstrated clinical utility as a diagnostic screening tool for hip fractures, particularly in facilities with limited radiological expertise. However, femoral neck fracture classification requires further refinement through dataset augmentation and advanced training methodologies to enhance detection accuracy for this radiologically challenging entity.