Mandible-YOLO: The fracture region is detected only once

Tao Zhou, Hongwei Wang, Kaixiong Chen, Zhe Zhang, Wenwen Chai, Huiling Lu

Abstract: BackgroundThe mandible is the part of human maxillofacial region where fractures often occur. In CT image, mandibular fractures are characterized by bone microfractures, impacted fracture and bone fractures due to different external power and angle. Therefore, the fracture region size and shape are different, and it is difficult to locate and recognize.MethodIn order to solve the above problems, the Mandible-YOLO model is proposed in this paper. The innovative works are as following: Firstly, the multi-scale residual feature enhancement module (MRFEM) is designed. In this module, convolution kernels of different scales are used to extract the features of multiple perceptive fields, which improve the perception ability of the model to different fracture region. Secondly, the spatial-channel feature hybrid module (SCFHM) is designed. This module combines spatial attention and channel attention of location information. Among them, channel attention is a frequency-domain transformed attention. The model recognition ability is improved for fracture region. Finally, the global–local feature hybrid module (GLFHM) is designed. This module introduces the Transformer mechanism into the C3 module. The model localization ability is improved for fracture region.ResultsThe mandibular fracture CT dataset is used to verify the model effectiveness, and comparing with the SOTA object detection model, the Mandible-YOLO model shows excellent performance in detection. The model achieved 97.02%, 97.12%, 93.82% and 95.11% respectively on mAP 0.5, Pre, Rec and F1 indexes.ConclusionThis model can improve the efficiency of mandibular fracture detection and has positive significance for mandibular fracture detection.

External IDs:dblp:journals/bspc/ZhouWCZCL25