面向下颌骨骨折检测的主辅网络YOLOv5模型

YOLOv5 model of primary and auxiliary networks for mandibular fracture detection

为了检测出下颌骨的骨折部位,帮助医生采取准确的治疗方案,同时解决骨折部位的CT影像受到出血和其余未骨折部位等的干扰所导致的骨折部位信息提取不充分、骨折部位检测精度不高等问题,提出了一种面向下颌骨骨折检测的主辅架构YOLOv5(primary and auxiliary YOLOv5,PA-YOLOv5)网络。首先,主干网络实现局部注意力和特征图多尺度通道混洗的耦合,获得影像的局部多尺度信息;其次,辅助网络实现结构重参数化和Transformer的耦合,获得影像的全局多尺度信息;最后,在所提出的数据集上进行消融实验和对比实验。结果表明,所提出的PA-YOLOv5网络的mAP(0.50)为98.18%,较原始YOLOv5网络提升了2.57%。所提出的网络能够较好地进行下颌骨骨折部位的检测,为医生提供相应的参考,从而辅助医生针对不同的骨折部位采取不同的治疗手段,为下颌骨骨折治疗提供一种新的诊断方法。

In order to solve the problems of insufficient feature extraction and low detection accuracy of the fracture site due to the interference by bleeding and other unfractured sites in fracture CT image,get a accurate detection of the fracture site of mandible and help doctors deciding treatment program,a primary and auxiliary YOLOv5(PA-YOLOv5)network was proposed for mandibular fracture detection in this paper.Firstly,the backbone network realized the coupling of local attention and multi-scale channel shuffling of feature maps to obtain the local multi-scale information of the images.Secondly,the auxiliary network realized the coupling of structure reparameterization and Transformer to obtain the global multi-scale information of the images.Finally,the ablation experiments and comparative experiments were carried out based on the proposed dataset.The experimental results show that the mAP(0.50)of the PA-YOLOv5 network proposed in this paper is 98.18%,which is 2.57%higher than the mAP(0.50)of the original YOLOv5 network.The network proposed in this paper can achieve better detection of the mandibular fracture site and provide corresponding reference for doctors,paving an effective way for taking different treatment methods for different fracture sites and provide a new diagnostic method for the treatment of mandibular fracture.