Research On Key Technology Of Intelligent Assistant Tongue Diagnosis

TCM tongue diagnosis is one of the four important diagnostic methods of observation,olfaction,inquiry,and palpation.Because of its intuitive,stable and easy to observe,it is more objective than other TCM diagnostic methods.To better inherit and develop traditional Chinese medicine,our country is actively promoting the modernization,objectification,and standardization of traditional Chinese medicine theory and diagnosis and treatment of traditional Chinese medicine.How to excavate the classic exposition and clinical experience of tongue diagnosis by ancient and modern medical experts,automate the traditional Chinese medicine tongue diagnosis,and exert its higher clinical application value is a major topic to promote the advantages of Chinese medicine.With the continuous development of medical image processing technology and artificial intelligence technology,combining with the clinical experience of traditional Chinese medicine experts to realize the quantification of tongue features and automation of tongue diagnosis are the mainstream research directions of modernization of traditional Chinese medicine tongue diagnosis.Although many researchers have done a lot of work on tongue segmentation and tongue feature analysis of two key technologies of tongue diagnosis automation,there are still many problems:for tongue segmentation,most studies use traditional image processing techniques,so the automatic segmentation cannot be achieved,and the accuracy of the segmentation is not high.For the feature analysis of the tongue,mainly used the classic machine learning method,and the data features need to be manually extracted,which cannot achieve end-to-end learning.Most research studied a single tongue feature,and the correlation between multiple features was ignored.As for TCM syndromes,due to the need for the participation of TCM experts,so there are few studies in this area.Because of the shortcomings of the above studies,this paper proposes a new tongue segmentation method and tongue image feature analysis method in combination with existing related technologies,and predicts the nine physiques of traditional Chinese medicine,providing a further research basis for the automation of traditional Chinese medicine tongue diagnosis.The main research contents of this article are as follows:The first,in the study of tongue image segmentation,this paper first proposes a new single-pixel loss function based on region correlation.The purpose is to guide the model to learn the correlation between pixels in a local region.The loss function can effectively combine the color and semantic relationship between adjacent pixels and guide the model learning based on the semantic information of the target pixel.Then,To solve the hole and glitch phenomenon of the segmentation result,this paper proposes a semantic segmentation model combining FCN and CRF,and experiments verify that CRF can effectively make the segmentation result smoother.Finally,the MIo U index of the FCN and CRF fusion model based on the new loss function is 96.55%.Secondly,based on the geometric features of the tongue,this paper proposes a tongue image geometric feature analysis model that fuses the spatial transformation layer and the VGG16 model.The spatial transformation layer is used to improve the effectiveness of the model for spatial invariance,and a fusion is also proposed to learn the residual feature coding layer and the texture feature recognition model of the VGG16 model,and transform the ordered features obtained by the convolution into an unordered representation,which can more effectively express the texture information of the tongue image.To make the model converge faster,the knowledge learned by the tongue image segmentation model is used to transfer the parameters of the above two models.Experiments have verified the effectiveness of the spatial transformation layer to improve the model invariance and the learnable residual coding layer to the tongue.The validity of the semantic representation of image texture features.Finally,the average recognition accuracy of all features reached over 82%.Third,this paper proposes a TCM physique identification model combining the tongue image segmentation model,tongue image analysis model,and decision tree.After effective tongue image segmentation and feature analysis,the obtained tongue image features are combined with physical features and age characteristics use a decision tree model to predict the TCM constitution.The experimental results show that the accuracy of the test results on the test set reaches 90%.At the same time,in order to verify the validity of the physical fitness identification model,this paper studies the consistency with the physical fitness identification method of the questionnaire.The prediction results on various constitutions are relatively consistent,and the overall consistency has reached76.7%.