Computer Aided Facial Bone Fracture Diagnosis (CA-FBFD) System Based on Object Detection Model

Abstract

Facial bone fractures must be diagnosed and treated as early as possible to avoid complications and sequelae. CT images need to be analyzed to detect fractures, but the analysis is time-consuming, and enough specialists are not available to analyze them. Many classification and object detection studies are being conducted to address these issues. The ability of classification-based studies to pinpoint the exact location of fractures is limited. Object detection-based research, by contrast, is problematic because the shape of a fracture is ambiguous. We propose a computer-aided facial bone fracture diagnosis (CA-FBFD) system to address the aforementioned challenges. This system adopts the object detection model YoloX-S, which is trained using only IoU Loss for box prediction, along with CT image Mixup data augmentation. For training, we used only nasal bone fracture data, whereas for testing, we used several other facial fracture data. During evaluation, the CA-FBFD system achieved an average precision of 69.8% for facial fractures, which is better than the baseline YoloX-S model by a large margin of 10.2%. In addition, the CA-FBFD system achieved a sensitivity/person of 100% for facial fractures, which is considerably better than that exhibited by the baseline YoloX-S model by a margin of 66.7%. Therefore, the CA-FBFD system can effectively minimize the labor of doctors who need to determine facial bone fractures in facial CT.