Research On Semantic Segmentation Algorithm For Stomatology Images

Panoramic X-ray and Cone beam CT are useful exams to complement dental analysis because of their cost-effectiveness and relatively low dose.Tooth segmentation is a crucial step before the diagnosis of dental diseases such as caries tooth extraction or endodontic diseases,but the manual labelling process is intensive and time-consuming.The primary challenges for tooth segmentation on dental images mainly lie in three aspects.(1)Large appearance variation:the appearance of main tooth regions may change dramatically for the cases with or without missing teeth,restortaion and dental appliance.(2)Local over luminance:dental materials used for restorations and prostheses result in bright metallic artifacts.(3)Vanishing boundary around the root regions:periodontal ligaments between roots and alveolar bones are thin and low-contrast in panoramic X-ray images.Therefore,this paper aims to solve the difficulties above and automatically segment teeth both in panoramic X-ray and Cone beam CT images.Detailed research content is as follows:A reasonable and interpretable data preprocessing algorithm is proposed according to tooth characteristics.This algorithm enhances the shape and gray features of teeth and it can be applied in both panoramic X-ray images and Cone beam CT images.According to the baseline experiment,it is found that the existing network still has some shortcomings in processing the semantic segmentation of teeth,such as the tooth boundary recognition is not precise enough,and the network prediction results are sensitive to noise.In view of the above problems,an automatic pixel-level tooth segmentation method ToothPix is designed based on conditional generative adversarial network structure.This methodology is the first to apply conditional GAN network on tooth segmentation task.The encoder-decoder based generator of ToothPix can contrast and learn grayscale and texture features of teeth guided by the discriminator.Experimental evidence on the panoramic X-ray dental dataset shows that the reasonable patch generation step can enhance the performance metrics of ToothPix.With patch size 1000×1000,Toothpix brings about a state-of-the-art segmentation performance on the LNDb panoramic X-ray dataset.In the end,this paper studies the scalability of the conditional confrontation generative model ToothPix in the application of 3D CBCT data.Specifically,first,the data set of cone-beam CT images of teeth currently being annotated is given.The data set is provided by the University of Electronic Science and Technology Hospital and has signed a data privacy protection agreement.This will be the first open source dental CBCT dataset.Secondly,the possibility of Toothpix processing 3D CT data is explored.The experimental results show that the extended Toothpix has strong robustness to tooth segmentation in cone beam CT images,and can effectively mine the inter-layer relationship and tooth features between the data.