Ultrasound Classficaiton Of Multi-Modal And Multi-Category Thyroid Nodules

Object: To establish and compare different predictive models of thyroid nodules based on logistic regression and decision tree,different indicators being included and nodular size being differentiated.Methods: Ultrasonic features of 1906 thyroid nodules in 1761 patients who had undergone sonography and fine-needle aspiration(FNA)or thyroid surgery in our hospital were retrospectively analyzed.According to the nodular size and whether color flow or elastographic indicators being included or not,the nodules were divided into 12 groups.By univariate and multivariate analysis,the significant sonographic features to diagnose thyroid nodules of each group were screened and compared,the logistic regression models were built,and the best cut-off value for benign and malignant prediction were calculated.Decision tree were used to classify 12 groups of nodules and establish diagnosis flow for thyroid nodules.The diagnostic performance of newly established models were validated and compared.Result: Significant features were used to diagnose nodules in all logistic regression models were hypoechoic,irregular margin,and microcalcification.Predominantly solid was an important indicator for the differentiation of benign and malignant thyroid nodules of large size.Taller-than-wide shape was a significant indicator of small malignant nodules of small size.The strain elastosonography character did show diagnostic value.The diagnostic performance of logistic regression models for malignant risk prediction were high,and the best predictive model is model 7,but the diagnostic performance was significantly reduced when the models performed bivariate prediction.Although the decision trees can provide seemly simpler diagnostic processes,the diagnostic performance of them were universal low and lower than the regression models.Conclusion: The most valuable indicators of malignant thyroid nodules are hypoechoic,irregular margin and microcalcification.New predictive models can fit in different nodular size and ultrasonic features included in scanning to distinguish malignant nodules from benign one.