Research On Prediction Of Benign And Malignant Thyroid Nodules Based On Dynamic Combination Of Multiple Classifiers

Thyroid nodule is a common disease,the causes of the disease are complex and variable,and the incidence rate is increasing year by year.At present,the diagnostic methods of thyroid nodules are mainly ultrasonic image diagnosis or chemical diagnosis.However,due to the complex and variable physical characteristics of thyroid nodules,as well as the external factors such as doctors’experience and medical equipment,the accuracy of diagnosis results is difficult to guarantee.Therefore,based on the physical characteristics of thyroid nodules collected by the hospital and some methods of machine learning,a scientific and intuitive method for predicting benign and malignant thyroid nodules is established in this paper,which can provide an auxiliary means for doctors’diagnosis.Combination learning is a new method in machine learning.It is widely used in various fields.In general,combination learning can achieve better performance than a single classifier.Therefore,this paper introduces combination learning into the prediction of benign and malignant thyroid nodules,and establishes an intuitive predictive model of benign and malignant thyroid nodules based on dynamic combination methods.This paper mainly uses the dynamic combination of multiple classifiers based on k-means clustering to analyze the data of patients with thyroid nodules.First,the thyroid nodule data is preprocessed,and the missing data of the discrete variables are filled by the k-nearest neighbor method;the missing data of the continuous variables are filled with the median.For the filled data,introduce interactive items in the data,and then select variables.Then,the base classifier is trained by C4.5,Naive Bayes(NB)and K-Nearest Neighbor(KNN)algorithm.On this basis,clustering method is used to divide the sample capability region,and then the multi-classifier dynamic combination model is obtained.Then,the model was evaluated with accuracy,ROC curve,AUC area and other evaluation indexes.Finally,test the selected model.The empirical analysis showed that the main factors leading to malignant thyroid nodules include:echo ratio,size,aspect ratio,morphology,marginal angle,internal structure,calcification,sonic,and the interaction of size and longitudinal diameter,longitudinal diameter and shape,longitudinal diameter and blood supply.Based on the selected factors,the classification results show that the classifiers trained by these variables achieve better prediction results than the classifiers trained with all variables,and the multi-classifier dynamic combination method based on clustering is better than the single classification method.