Deep Learning-Based Segmentation Of Tooth-Marked Tongue Images

Tongue diagnosis is an important part of TCM inspection.The features of tooth marks in tongue images refer to the marks of teeth that can be seen on the edge of the tongue.Tooth-marked tongue is an important objective indicator for diagnosing spleen deficiency syndrome,and identifying this indicator can promote the differentiation of symptoms and treatment selection in TCM.Therefore,some researchers try to finely segment the tongue from the tongue image,so as to serve the automatic recognition research of TCM tongue diagnosis.However,segmentation of tooth-marked tongue images is extremely challenging.Tooth marks can be divided into mild,moderate and severe.Different degrees of tooth marks correspond to different degrees of spleen deficiency and wet weight,which are reflected in subtle differences on the edge of the tongue.Therefore,the high-quality segmentation of the tongue body boundary of the tooth-marked tongue is the key to realize the distinction between the normal tongue and the tooth-marked tongue,and on this basis,the different marked degrees can be further accurately distinguished.In this paper,two segmentation methods are proposed for the segmentation task of tooth marks and tongue images.The research contents are as follows:Firstly,a tooth-marked tongue image segmentation method based on rough prediction guided residual U-network is proposed,which realizes a lightweight framework.In the first stage,a full convolution network architecture composed of Res Net basic blocks is constructed to realize the initial segmentation of tooth-marked tongue image;In the second stage,guided by the obtained rough segmentation results,the residual U-shaped structure is constructed,and the difference between the initial segmentation image and the original image is learned to obtain the finer prediction results of the boundary,so as to realize the refinement of the segmentation boundary.In addition,based on the strategy of in-depth supervision,the model designs and deploys a new loss function to accelerate the convergence of in-depth network model.The experimental results show that this method is applied to the tooth-marked tongue image segmentation,which improves the average 95 HD to 14.29 pixels and the maximum 95 HD to 32.02 pixels,and achieves excellent segmentation performance.Then,aiming at the possible problems in the use of atrous convolution stacking,a small cell module of hole residual is proposed.The first stage network framework of tooth mark tongue image segmentation method based on rough prediction guided residual u-network is improved.Specifically,the output characteristic map of the fourth part of resnet-50 backbone network is used as the input of the module,and four atrous residual small units with atrous rates of [2,3,4,5] are connected in series to realize feature enhancement,so as to obtain a better initial segmentation map.Secondly,the thinning of the initial segmentation image is consistent with the previous work,and the final two-stage tooth-marked tongue image segmentation method fused with cavity convolution is obtained.The experimental results show that the model improves the performance of the basic network in the segmentation task.It can be seen that the Miou reaches 97.02%,the MPA reaches 98.45%,the average 95 HD is increased to 10.35 pixels,and the maximum 95 HD is increased to 22.14 pixels.In addition,this paper also realizes the lightweight network in the case of high-level segmentation performance.