Research On Tooth Segmentation Model Based On Deep Learning

Using computer-aided dental segmentation technology for the diagnosis of dental lesions in the human body,location,extraction,and quantitative analysis is the prerequisite and guarantee for maximizing the application value of dental images.In medical image processing,dental image segmentation is one of the important means for assisting medical personnel in judgment.Currently,medical images used in dentistry are mainly divided into two types: panoramic X-rays and CBCT images.This article builds a deep learning-based dental segmentation system around the problem of tooth segmentation,which segments both panoramic X-rays and CBCT images of teeth.The research content includes the following aspects:(1)In response to the problem that the tooth boundary is fuzzy,the contrast between teeth and alveolar bone is low,and accurate segmentation of teeth is difficult in panoramic X-rays,this article proposes a Teeth Seg Net model for dental panoramic X-rays with context semantics and enhanced tooth boundary.A multi-scale aggregation attention block MAB is designed in the bottleneck layer,which can effectively extract tooth shape features and adaptively fuse multi-scale features.In addition,a designed extension mixed self-attention block DHAB effectively suppresses background information irrelevant to the task without increasing network parameters.Compared with multiple networks,Teeth Seg Net has the best segmentation effect,and the Dice coefficient can reach 0.9439.(2)In response to the problem that automatic segmentation of teeth on CBCT images is inaccurate due to severe artifacts,traditional U-Net methods are not effective.In this paper,a3.5D U-Net is proposed,which incorporates majority voting mechanism and erosion and dilation methods to improve its performance in tooth segmentation on CBCT images.Compared with 2D U-Net and 2.5D U-Net,the 3.5D U-Net proposed in this paper achieves the best performance in terms of DSC,Ac,Sn,Sp,PPV,and NPV indicators.(3)A new weakly supervised tooth segmentation model based on a combination of a deep learning-based target detection method and a level set method is proposed to address the problem of the large cost of CBCT image standards and the inevitable admixture of human noise.The model automatically generates pseudo-annotation information of dental data for tooth segmentation by combining an active contour model and a target detection model,and the method achieves a reduction in the workload of data annotation.Experimental results show that even with interference from gingival tissue and false teeth,the model can accurately segment teeth in CBCT images.The average of Dice coefficient,Jaccard coefficient,BF score,and precision is 0.9477,0.9031,0.9789,and 0.9479,respectively.(4)A dental segmentation platform incorporating the above three dental segmentation models is built to enable medical professionals to select different segmentation models to segment dental images according to the actual situation.