Quantification And Classification Of Cracked Tongue Based On Machine Learning

Cracked tongue is a vital information source of TCM tongue diagnosis and treatment,which can objectively as well as accurately reflect the changes of some typical diseases and TCM syndromes.The traditional diagnosis of cracked tongue is easily affected by doctor’s experience,environmental changes and other factors,especially the lack of objective quantification of the severity of cracked tongue.With the help of classical machine learning algorithm,this paper applies digital image processing and pattern recognition technology to cracked tongue intelligent diagnosis,in order to improve the objectivity,quantification and standardization of traditional diagnosis.This paper constructs a complete evaluation framework to distinguish the severity of tongue cracks.Firstly,a four step accurate crack extraction method is designed,including using“global truncation threshold segmentation" to distinguish crack and non-crack areas,using "black-hat transformation" to improve the clarity of crack areas,reducing pseudo crack noise through "global threshold",and using "binary small object removal method" to eliminate pseudo cracks.Then,according to the area ratio and average gray ratio of crack and non-crack area in tongue image,the tongue crack visibility index(CVI)and depth index(CDI)are defined.On this basis,the tongue crack severity index(CSI)is further defined.Taking the doctor’s mark as the gold standard,the quantitative standard of tongue crack severity is given,if CSI ≤0.2,it belong to mild crack and if CSI>0.2,it belong to severe crack.Finally,three kinds of features of tongue image are extracted with the help of calculation tools,including histograms of oriented gradients feature,color feature and texture feature.Four classical classifiers are input by different feature combination methods,and the gold standard and quantitative standard are used as labels to identify mild or severe tongue’s cracks respectively.The classification results of 95 crack tongue images show that the prediction results of each classifier under the two standards are no significant difference,which verifies the consistency between the quantitative standard and the gold standard.In addition,support vector machine achieves the best prediction performance under the combination of“color+texture",which shows that the evaluation framework proposed in this paper can effectively distinguish the severity of cracked tongue.In order to further verify the applicability of the extracted three types of tongue image features and the commonly used machine learning classifiers to the intelligent recognition of cracked tongue,the different combinations of the three types of features are input into four classical classifiers to recognize 2033 tongue images taken by smart phones.The experimental results show that different features(combinations)have different recognition effects in different classifiers.Multi-feature fusion often contribute to improve the recognition ability of classifiers.Compared with the basic classifier,the ensemble classifier has higher robustness and generalization.In particular,the AdaBoost classifier integrating three kinds of features has achieved recognition effect almost comparable to that of deep learning model(residual neural network and densely connected convolutional networks).