A Key Technique For Risk Assessment Of Surgical Complications Of Inferior Alveolar Nerve Injury Based On Deep Convolutional Networks

Removal of mandibular wisdom teeth is one of the most common procedures in oral and maxillofacial surgery.To avoid complications such as inferior alveolar nerve canal injury,the risk of inferior alveolar nerve canal injury must be assessed preoperatively.The preoperative risk assessment is divided into three parts.First,the anatomical features of the patient’s X-ray cephalometric images are localized,and the patient’s craniomaxillofacial anomalies are evaluated by analyzing the statistical information composed of these features.Third,the number of roots of the patient’s mandibular wisdom teeth was obtained from panoramic radiography.The accuracy of the positioning and judgment in these three steps directly affects the correctness of the risk assessment.At present,these three steps are mainly performed manually in the actual dental clinic.The manual assessment is time-consuming even for doctors with extensive clinical experience,and can be seriously affected by subjective factors such as the experience and energy of the observer.Therefore,the development of automatic and high-precision key techniques for risk assessment of surgical complications of inferior alveolar nerve injury has important theoretical research value and clinical significance.To address the problem of automatic localization of anatomical feature points in X-ray cephalometric images,this paper proposes an algorithm for automatic localization of anatomical feature points in multiscale two-dimensional X-ray cephalometric images based on relative position constraints.The algorithm uses UNet2+ network as a framework to build a depth model for automatic localization of anatomical feature points.The algorithm model has multiple encoding-decoding modules of different scales and cascades each module by jumping connections,while the weights of each module are optimized by using depth supervision.In the training phase,the topology consisting of the absolute positions of anatomical feature points and the relative positions between feature points is optimized by using the joint loss of Mean Squared Error Loss(MSELoss)and Position Constraint Loss(PCLoss).The test phase obtains the final predicted feature points by decoding the heat map and linear transformation mapping.The experimental results show that the method can effectively learn the topological spatial relationship and the absolute position distribution of each anatomical feature point,and thus can predict the anatomical feature point coordinates with a high accuracy.For determining the spatial relationship between the mandibular obstructed wisdom teeth and the inferior alveolar nerve canal on the panoramic radiography,this paper then proposes a multiresolution target detection network-based algorithm for risk assessment of inferior alveolar nerve injury.The algorithm is based on the YOLOv5 framework.The algorithm model consists of a basic skeleton end,a detection end and an output end,where the input image is extracted from the skeleton end and fed into the detection end,and the detection end is divided into multiple scales by a pyramidal structure for detecting targets of different sizes in the output end.The overall flow of the algorithm is a data processing phase,a training phase and a testing phase.In the data processing phase,data enhancement is performed on the panoramic radiography while using the spatial relationships determined on the Cone-Beam Computed Tomography(CBCT)images as the gold standard.In the training phase,detection frame position loss,confidence loss,and classification loss are used as joint losses to optimize the model,while a hyperparameter search is used to select the optimal set of hyperparameters for model training.In the testing phase,Test Time Augmentation(TTA)is used to improve the final prediction accuracy.The experimental results show that the method can effectively resolve the anatomical information of mandibular obstructed wisdom teeth and inferior alveolar nerve canal in the panoramic radiography by combining multi-scale objectives,and thus can predict the spatial relationship between them with high accuracy.In order to determine the number of wisdom tooth roots of mandibular blocked wisdom teeth from the panoramic radiography,this paper proposes an automatic algorithm for determining the number of mandibular blocked wisdom tooth roots based on multi-resolution target detection network based on the above work.The algorithm uses the number of roots of CBCT images as the gold standard for labeling,optimizes the parameters by means of hyperparametric search,trains the multiresolution target detection network model to establish a nonlinear relationship between image features and the number of roots of mandibular obstructed wisdom teeth,and uses this relationship to automatically predict the number of roots of mandibular obstructed wisdom teeth from the newly obtained surface tomographic images,and also applies the TTA algorithm to enhance the final prediction accuracy.The experimental results show that the method can effectively resolve the anatomical information of mandibular obstructed wisdom teeth in the surface tomographic images by combining multi-scale objectives,and thus can predict the number of roots of wisdom teeth with high accuracy.