| The extraction of mandibular impacted third molars is a common operation in the clinical clinic of oral and maxillofacial surgery,which is prone to surgical risks such as nerve injury in the mandibular canal.The factors related to surgical risks are the position relationship between the root of the third molar and the mandibular canal,the external resorption of the root of the second molar,the impacted type of the third molar,the impacted depth of the third molar and so on.Among them,the relationship between the root of the third molar and the mandibular canal and the external resorption of the root of the second molar are the more important factors.To accurately judge the relationship between the root of the impacted mandibular third molar and the mandibular canal from the medical images before operation and whether there is external resorption of the root of the mandibular second molar is an important step in the risk assessment of mandibular impacted third molar extraction.At present,oral clinic still mainly adopts artificial methods to judge the relationship between the root of the impacted mandibular third molar and the mandibular canal from the panoramic radiographs and whether there is external resorption of the root of the second molar,the accuracy and diagnostic efficiency of risk assessment of impacted mandibular third molar extraction are greatly influenced by doctors’ experience,energy and other subjective factors.Therefore,it is of important clinical value to develop an automatic detection method for the relationship between the root of the impacted mandibular third molar and the mandibular canal and the external root resorption of the mandibular second molar.In order to improve the accuracy and efficiency of identifying the position relationship between the root of impacted mandibular third molar and mandibular canal in panoramic radiographs,an automatic detection method based on single-step neural network was proposed in this paper.In this method,the automatic detection of the relationship between the root of the impacted mandibular third molar and the mandibular canal was regarded as a combination of regression task and classification task,and no manual operation was involved in the testing process.Based on the YOLOv5 network,a depth convolution neural network was constructed,which can completed the task of classification and location at the same time.The spatial position relation information obtained from the corresponding cone beam CT(Cone-beam Computed Tomography,CBCT)image was used as the classification gold standard to train it to learn the nonlinear relationship between the features of the panoramic radiographs and the root of the third molar contacting the mandibular canal.After the newly obtained panoramic radiograph was input into the trained network model,the probability of the contact between the root of the impacted mandibular third molar and the mandibular canal can be obtained,and the area of contact between the root and the mandibular canal can be predicted at the same time.The experimental results show that,compared with other methods,this method can improve the accuracy of judging the relationship between the root of the impacted mandibular third molar and the mandibular canal,and can predict the area where the root of the impacted mandibular third molar is in contact with the mandibular canal.In order to reduce the complexity of the algorithm and improve the detection accuracy,the feature extraction network was improved on the basis of the work in the previous chapter.An automatic detection method of the relationship between the root of impacted mandibular third molar and mandibular canal based on improved single-step neural network was proposed.In this method,the Ghost Bottleneck module and SELayer(Squeeze-and-Excitation Layers)module were used to replace the CSP(Cross Stage Partial)module in the depth model feature extraction network used in the previous chapter,so as to reduce the computational complexity and improve the efficiency of feature expression.The experimental results show that compared with other network models,the improved single-step neural network can effectively reduce the network complexity and further improve the detection accuracy of the relationship between the root of the impacted mandibular third molar and the mandibular canal in the panoramic radiographs.In addition,this paper proposed an automatic detection method of external root resorption of mandibular second molars in panoramic radiographs based on YOLOv5 frame.The depth neural network was constructed based on the YOLOv5 model,which can simultaneously predicted whether there is external root resorption in the panoramic radiographs of the mandibular second molars and the location of the focus of external root resorption.No manual operation was involved in the prediction process.The experimental results show that this method can effectively judge whether there is external resorption of the root of the mandibular second molar in the panoramic radiographs and detect the area of external root resorption. |
