A COMPARISON OF A DEEP LEARNING NEURAL NETWORKS MODEL WITH HUMAN OBSERVATIONS FOR DETECTING NON-OBVIOUS RADIUS AND CARPAL BONE FRACTURES

Introduction: Patients with hand and wrist trauma are frequently diagnosed in the emergency department. Deep learning algorithms could potentially become powerful tools to diagnose fractures from X-ray wrist images. This study aims to assess the diagnostic performance of a deep learning algorithm in detecting wrist fractures that are difficult to detect through radiographs. Methods: This retrospective study included adult patients with hand/wrist trauma who undergo CT imaging. CT imaging of injured areas, interpreted by an expert radiologist were considered as “ground truth” (GT). There were 313 cases, a total of 121 fractures (82 radius, 39 carpal bones) were identified as GT from CT images. Using the algorithm, fracture detection procedure was performed on dataset of hand and wrist X-ray images. The same datasets were evaluated by four emergency medicine doctors. Diagnostic performances such as accuracy, area under curve, sensitivity, precision and F1 score were calculated. Agreement (Kappa coefficient (κ)) between GT, observers and deep learning algorithm was determined. Results: The algorithm showed 69.6% accuracy, 57% sensitivity and 61.6% precision. Emergency medicine doctors showed better diagnostic performance with higher accuracy, sensitivity and precision and AUC values. The interobserver agreement among four EM doctors was moderate whereas the agreement with the algorithm was only fair. Conclusions: The Deep learning algorithm demonstrated an accurate detection of fractures in wrist X-rays and it had capabilities that were comparable to those of emergency medicine physicians, but the algorithm mentioned needs to be further improved to produce better outcome.