Abstract: Objective To establish a combined model based on ultrasound radiomics combined with clinical parameters and conventional ultrasound features, and to verify its value in predicting benign and malignant thyroid nodules. 
Methods A total of 209 patients with 237 thyroid nodules confirmed by pathology were retrospectively included and randomly divided into training set (190 thyroid nodules) and validation set (47 thyroid nodules) according to an 8∶2 ratio. Pyradiomics was used to extract radiomics features, least absolute shrinkage and selection operator (LASSO) was used for feature screening, and k-Nearest Neighbor (KNN) machine learning algorithm was used to construct radiomics model. Univariate and multivariate Logistic regression were used to screen clinical parameters and conventional ultrasound features. The risk factors analyzed by multivariate Logistic regression were included in the k-nearest neighbor machine learning algorithm to construct a clinical prediction model. A combined model was constructed using the screened radiomics features and the risk factors analyzed by multivariate Logistic regression. The diagnostic efficacy of each model was compared by receiver operating characteristic (ROC) curve. 
Results Multivariate Logistic regression results showed that microcalcification, vertical position and age were independent risk factors for differentiating benign and malignant thyroid nodules (OR=6.082,13.761,0.938, all P＜0.001). The clinical prediction model was constructed using the above risk factors, and the area under the curve (AUC) of the training set and the validation set was 0.881 and 0.704, respectively. The ultrasound radiomics model was established by using the six key radiomics features screened by LASSO algorithm. The AUC of the training set and the test/validation set was 0.830 and 0.819, respectively. The risk factors analyzed by multivariate Logistic regression and radiomics score were used to construct a combined diagnostic model. The AUC of the training set and the validation set was 0.939 and 0.854, respectively. The diagnostic efficiency was better than that of a single clinical prediction model or ultrasound radiomics model. The decision curve showed that the combined model had good clinical application value. The calibration curve showed good fit between the combined model and the ideal model. 
Conclusion The combined model constructed by radiomics based on contrast-enhanced ultrasound combined with clinical parameters and ultrasound features can effectively identify benign and malignant thyroid nodules. 

Key words: thyroid nodule, radiomics, nomogram, ultrasound, artificial intelligence