Research On Application Of Traditional Chinese Medicine Tongue Images Classification Based On CNN

The tongue image diagnosis is an important part of Traditional Chinese Medicine(TCM).The doctors can learn about the patient's physical condition and changes by looking at the patient's tongue.Among many physical symptoms reflected in the tongue image,the syndromes in TCM theory are an important state of the patient's body revealed and reflected in the tongue diagnosis.The “cold” and “hot” are two typical examples of the human body condition.These two kinds of syndromes is also the objective symptom of TCM tongue diagnosis in this paper.However,the tongue diagnosis process is affected by other factors such as illumination,shooting angle and the doctor's subjectivity,making it difficult for the objective research to achieve rapid breakthrough.In recent years,with the deep research of machine learning in the field of image recognition,these algorithms are widely used in image classification problems.Subsequently,through the combination of computer image processing technology and medical imaging,people are increasingly approaching the goal of computer-assisted diagnosis.This exploratory study has brought new directions and opportunities for the objective study of tongue diagnosis in TCM.The researchers made the tongue diagnosis standardized and objective by acquiring tongue image data,image preprocessing,machine learning,and constructing classification models.However,a key step in the use of traditional machine learning methods is the feature extraction.But this step requires the researchers to have a large amount of domain knowledge and experience,at the same time their subjective awareness in the extraction step may easily lead to blindness in feature extraction,so that the classification result cannot be optimal.This paper studied the convolutional neural network structure in deep learning,automatically extracts feature information,and explored the feasibility of the method in TCM tongue classification model,and at the same time assisted in the further realization of TCM tongue image diagnosis objectivization.This paper studied the progress of tongue image objectivization and convolutional neural network and carried out the following work on the basis of it: First,during the study period,this paper,through cooperation with Shanghai University of TraditionalChinese Medicine,was divided into two original collections.The data were 398 cases,of which proportion of normal : hot syndromes : cold syndromes was 79:149:170.Data preprocessing is performed on the data set,including data selection and tongue image segmentation.Secondly,through the research and experiments,this paper selected and implemented four convolutional neural network structures—Conv4,AlexNet,ResNet18,and ResNet50—from the perspective of network model and experimental results.Third,this paper designed and simulated two-class experiments,including normal-syndrome,normal-heat syndrome,normal-cold syndrome,and heat syndrome-cold syndrome,and continued to extend the three-category experiment which is normal-heat syndrome-cold.Forth,through training and cross-validation of the network model in the four experimental data,the accuracy curves of the corresponding training samples and test samples were obtained.By comparing the four network models and the traditional machine learning methods,the accuracy and fitting requirements were obtained.From the perspective of the number of rounds,the performance of each model on this type of problem was found.Among them,most convolutional neural network structures achieved very good experimental results,such as ResNet18 has achieved a accuracy with 94.4% in the normal-heat syndrome test,and a accuracy with 89% in the 3-classification syndrome test.