Application Of Machine Learning In Ultrasonic Image Recognition Of Malignant Thyroid Nodules And Tongue Classification In Traditional Chinese Medicine

Thyroid tumor is a common endocrine tumor.The incidence of thyroid malignant tumor increases year by year.Thyroid cancer has a lower malignancy and slower development,so the mortality rate is relatively low in malignant tumors.Early surgery is an important way to improve survival rate and prognosis.It is obvious that the early diagnosis of thyroid cancer is significant.Ultrasound is widely used in clinical as an efficient and noninvasive examination for the differentiation of benign and malignant thyroid nodules.Except western medical surgery,TCM therapy can effectively relieve symptoms and improve prognosis.However,whether it is ultrasound examination or TCM syndrome differentiation,all depend on the technical level and clinical experience of the diagnosing physician.It is subjective and has a certain rate of missed diagnosis and misdiagnosis.The application of machine learning can effectively solve this problem.Machine learning can effectively improve the accuracy and objectivity of diagnosis.And provide standardized and intelligent services for subsequent treatment.Objective:This paper analyzes the relationship between the ultrasound images and the benign and malignant thyroid nodules from two angles of TCM and western medicine,and the degree to which these features affect the diagnosis of benign and malignant nodules.To explore the distribution of TCM syndrome types and tongue image characteristics of thyroid cancer,and provide guidance for the treatment of syndrome differentiation.According to the features of ultrasound images,TCM syndromes,and tongue image that are closely related to malignant thyroid nodules,we create a classification model based on machine learning convolutional neural network,and evaluate its classification performance,so as to establish a thyroid nodule intelligent diagnosis and treatment system,which is including benign and malignant thyroid nodules and tongue image classification two subsystems.Methods:In this study,a total of 241 patients,who were diagnozed as thyroid nodules by pathological diagnosis,were collected in our hospital,including 171 benign nodules and 70 malignant nodules.These patients' ultrasound images,TCM syndromes,tongue image and pathological classification data were recorded in detail,and then analyzed these data by univariate analysis and multivariate logistic regression analysis,respectively.To explore the relationship between ultrasound image features and benign and malignant thyroid nodules,and how these features affect the diagnosis of benign and malignant.Try to find out the relationship between the TCM syndromes of malignant thyroid nodules and the characteristics of ultrasound images and pathological classification.To sum up the tongue features of patients with malignant thyroid nodules.Then,with the help of machine learning,we choose two entry points from western medicine and Chinese medicine,(1)these ultrasound features which were most closely related to the benign and malignant nodules were input into benign and malignant classification model as input values,(2)these tongue features of thyroid cancer patients were input into tongue image classification model as input values,and these two classification models were constructed based on the ResNet algorithm in the convolutional neural network structure.Used the ROC curve,accuracy,sensitivity,specificity,accuracy,etc.to evaluate the classification performance of the model.Results:1.In univariate analysis of the relationship between thyroid nodule ultrasound image performance and benign and malignant,the differences of five variable composition ratios are statistically significant in benign and malignant groups,such as nodular echo,edge,A/T,posterior attenuation,and calcification(P<0.05).the differences of nodule size,internal echo conditions,halo,blood flow,location,and lymph node composition ratio are not statistically significant between benign and malignant groups(P>0.05).2.In the multivariate logistic regression analysis of the relationship between thyroid nodule ultrasound image performance and benign and malignant,the regression coefficient P value(at the level of a=0.05)shows that ten indicators' contribution are not significant,including gender,age,size,echo,internal echo,posterior attenuation,halo,blood flow,position,and lymph nodes P>0.05,the differences of margin,A/T,calcification in the benign group and malignant group have statistics significant(P<0.05).After variable selection,logistic regression analysis was conducted again.The regression coefficient and OR value are compared to find that the results of A/T has the greatest impact on benign and malignant,followed by calcification and margin.3.Correlation analysis of TCM syndrome types,ultrasound manifestations,and pathological classification of malignant thyroid nodules shows that there are statistically significant differences in nodule echo,margin,and calcification,P<0.05.There are no statistically significant difference in size,internal echo,aspect ratio,posterior attenuation,halo,blood flow,location,lymph nodes,etc.,P>0.05.Thyroid nodules in liver depression and qi stagnation type are low echo,equal or high echo and mixed echo,most of them with marginal rules or less regular,and microcalcification or gross calcification,microcalcification is slightly more than gross calcification;Thyroid nodules in phlegm and blood stasis type are mostly low or extremely low echo,with irregular margin and microcalcification;Thyroid nodules in stasis and yin type are mostly low or extremely low echo,with regular or irregular margins,without calcification.There is no obvious correlation between TCM syndrome types and pathological types of malignant thyroid nodules,P value>0.05.4.The TCM syndrome types of malignant thyroid nodules are mainly liver qi stagnation,phlegm and blood stasis,stasis and heat stagnation,and the basic pathogenesis is qi stagnation,phlegm coagulation,and blood stasis.The characteristics of tongue images in patients with malignant thyroid nodules are mainly concentrated in dark red or dark purple tongues,with tongue marks or ecchymosis,and thin white coat on the tongue,which is consistent with its basic pathogenesis.5.1n the machine learning part,the CNN-ResNet50 model is used to classify the benign and malignant thyroid nodules in the verification set.The results show that:accuracy:96.08%,sensitivity:92.45%,specificity:96.30%,F1 value:94.23%,AUC:93.40%,the accuracy rate:94.39%.And on the test set,the trained CNN-ResNet50 model was compared with 6 ultrasound physicians for distinguishing benign and malignant nodules.The ResNet50 model's classification accuracy for benign and malignant nodules was 94.26%.The average classification accuracy of 6 professional ultrasound doctors for benign and malignant nodules was 90.98%,the difference is not statistically significant,P value>0.05.The same model was used to classify dentate tongue and ecchymotic tongue.The results show that:total accuracy:90.76%,total sensitivity:93.02%,total specificity:84.19%,total F1 value:91.88%,total accuracy rate:90.15%,AUC:96.00%.Conclusions:1.Ultrasound manifestations are related to benign and malignant thyroid nodules.The impact of ultrasound manifestations on benign and malignant in descending order is:A/T>calcification>margins,of which A/T?1 is the highest risk for thyroid malignant nodules.The distinction between benign and malignant contributes the most.2.TCM syndrome types of malignant thyroid nodules have a certain correlation with ultrasound performance.Ultrasound examination can be used as an extension of the TCM"inspection",providing an objective basis for syndrome typing and syndrome differentiation.3.CNN-ResNet50 model can accurately distinguish the benign and malignant thyroid nodules and tongue images of patients with malignant thyroid nodules.It shows that the model has strong ability of feature learning and extraction,and finally the ability to classify the nodule nature is comparable to that of professional ultrasound doctors.It proves that the CNN model has the ability to distinguish thyroid tumors from Western medicine examinations and TCM syndromes,and can more intelligently assist diagnosis,lay a solid foundation for the subsequent application of machine learning in the diagnosis and treatment of Chinese and Western medicine in thyroid tumors.