For artificial intelligence assistance in the detection of fracture sites， the fracture sites are usually accompanied by bleeding and other symptoms. Further， CT images taken in different positions have large differences， the size of fracture sites varies， and the bleeding sites and surrounding tissues may interfere with the detection of fracture sites， leading to insufficient feature extraction and the problem of low detection accuracy. Therefore， the 3M-YOLOv5 network is designed to detect mandibular fracture sites. First， the dense module is used in the feature extraction network to improve the feature extraction capability of the network by using the dense connection property. The local and global attention module （lgaM） is used to extract the global information of CT images. Second， a lightweight multiscale dense block （lmdM） is designed to extract the multiscale features of the fracture sites with fewer parameters. Third， a cross-dimension bidirectional feature fusion module （cdbfM） is designed in the feature enhancement network to make the height， width， and channel of the feature maps interact with each other， and trainable weights are introduced to balance the fusion importance of the feature maps with different scales. Finally， to verify the effectiveness of the proposed network， ablation and comparison experiments are conducted on a self-built dataset. The results show that when the confidence threshold is 0.5， the mAP value， F1 value， recall rate， and precision rate of the proposed network are 99.17%， 99.06%， 98.81%， and 99.32%， respectively. The proposed CT image detection network for mandibular fracture can better detect the fracture sites in the image than existing methods， which is a good reference for doctors to make a corresponding treatment plan based on the detection results.