The Application Value Of Different Machine Learning Models Based On CT Imaging Omics In Differential Diagnosis Of Benign And Malignant Thyroid Nodules

Objective: To explore the value of seven machine learning models based on CT imaging omics in distinguishing benign and malignant thyroid nodules,and to seek the best differential diagnosis model.Methods:(1)116 patients(60 benign and 56 malignant)with thyroid nodules diagnosed and treated in the people’s Hospital of Guangxi Zhuang Autonomous Region from January2019 to March 2022 were collected.(2)After preprocessing the CT plain scan and enhanced arterial phase images,import them into the 3D slicer software to manually draw the region of interest(ROI)layer by layer,and extract the image omics features using the pyrodiomics plug-in package of Python platform,with a total of 129 feature values.Firstly,two independent sample t-test or rank sum test were used to screen the statistically significant features,and then lasso(least absolute shrinkage and selection operator)was used to further reduce the dimension.Finally,15 and 11 best imaging omics features were obtained in plain scan group and enhanced arterial phase group respectively.(3)The patients were randomly divided into training group and test group according to the ratio of 7:3,the following seven machine learning models(MLM)were established to identify benign and malignant thyroid nodules: logistic regression model,random forest model,linear support vector machine model,polynomial support vector machine model,Gaussian support vector machine model,decision tree model K-nearest neighbor algorithm(KNN)model.Results: Forest model was the most effective in differential diagnosis of benign and malignant thyroid nodules in plain scan group and enhanced arterial phase group.(1)The sensitivity,specificity,accuracy and AUC values of forest model training group in plain scanning group were 0.92,0.97,0.95 and 1.00 respectively,and those of test group were 0.92,0.97,0.95 and 0.89 respectively.(2)The sensitivity,specificity,accuracy and AUC of the forest model training group in the enhanced arterial phase group were 0.94,0.93,0.93 and1.00 respectively,and the sensitivity,specificity,accuracy and AUC of the test group were0.97,0.95,0.96 and 0.95 respectively.Conclusion:(1)Radiomics machine learning model is a new method to distinguish benign and malignant thyroid nodules.(2)Forest model is the most effective method for differential diagnosis of benign and malignant thyroid nodules in plain CT scan group and enhanced arterial phase group.(3)CT enhanced arterial forest model is better than CT plain scan forest model in the differential diagnosis of benign and malignant thyroid nodules.