Locating Anatomical Landmarks On 2D X-ray Cephalometric Image By Using Generative Adversarial Network

Since it was introduced into the field of Orthodontics,X-ray cephalometric analysis technology has become an important tool for the study of craniofacial growth and development,clinical deformity diagnosis and treatment design.The X-ray cephalometric analysis should first determine the location of each anatomical landmark,then analyze the line angle relationship of these anatomical landmarks,and then draw the diagnosis conclusion after comparing with the normal person.Therefore,the positioning accuracy of anatomical landmarks directly affects the accuracy of cephalometric analysis.At present,the anatomical landmarks of cephalometric images are still marked manually.However,even experienced doctors’ manual marking is very time-consuming,and is subject to subjective factors such as doctors’ personal experience and energy.Therefore,the development of automatic location of anatomical landmarks in cephalometric images can be used as a solution to these problems,and it is of great scientific significance and clinical value.In order to solve the problem of automatic location of oral anatomical landmarks in two-dimensional cephalometric images,this paper proposes an algorithm of automatic location of oral anatomical landmarks in two-dimensional cephalometric images based on adversarial auto-encoder network.The whole process of the algorithm includes data preprocessing stage,training stage and testing stage.Firstly,in the data preprocessing stage,the 2D X-ray cephalometric image is clipped,and the corresponding distance map is calculated according to the anatomical feature points to be detected.Then in the training stage,the model uses the preprocessed cephalometric image and the corresponding distance map as the input data to train the generator,which outputs the predicted distance map,and then uses real and predicted distance maps to train the discriminator.Finally,in the test stage,the trained model can predict the distance map according to the input cephalometric image,and the predicted distance map can be calculated by regression voting method to obtain the coordinates of the feature points to be detected.In this paper,the method creatively uses the generation model to complete the prediction of the coordinates of the anatomical feature points.The experimental results show that the method can generate the distance map with high quality,and thus calculate the coordinates of the oral anatomical landmarks with high accuracy.At the same time,due to the problems of relatively fixed anatomical structure and insufficient semantic information in the X-ray cephalometric image,considering that the U-net network structure can effectively integrate the multi-scale features of medical images,this paper also proposes an automatic positioning algorithm of the two-dimensional X-ray cephalometric image oral anatomical landmarks based on the adversarial U-net network.The purpose of this model is to combine the advantages of U-net structure and generate adversarial network framework,and make use of U-net network’s unique U-shape structure and skip connection mechanism,to fuse the X-ray cephalometric images in different scales,and then combine the adversary idea of generate adversarial network,so as to improve the quality of predicted distance map and the accuracy of anatomical landmarks positioning.