The Value Of Us-based Radiomics In Differential Diagnosis Of Thyroid Cancer And Cervical Lymph Node Metastasis

Objective To comparatively assess the performance of US（ultrasound）-based radiomics and ACR-TIRADS（American College of Radiology Thyroid Imaging,Reporting and Data System）in identifying malignant nodules and predicting cervical nodules metastasis.Methods A total of 339 patients with thyroid nodule who were operated in the First Affiliated Hospital of Guangxi Medical University from March 1^st,2019 to September 15^th,2019 were analyzed retrospectively in this study.First,according to the characteristics of conventional ultrasound,thyroid nodules were classified by ACR-TI-RADS.US was also used to determine if there were lymph node metastasis.Then,ITK-SNAP was used to manually delineate the ROI（region of interest）of nodules in the largest US slice,and 5234 radiomics features were extracted from each ROI through Intelligence Foundry.Patients were divided into train set and test set by stratified sampling.Six feature selection methods and ten modeling classifiers were used for modeling,and then the test set was used to evaluate the model performance.Finally,the AUC（area under curve）was used to evaluate the effectiveness of US-based radiomics model,conventional ultrasound preoperative prediction of thyroid cancer and cervical lymph node metastasis.Results The AUC of ACR-TIRADS for thyroid nodules diagnosis was0.825.The AUC of conventional ultrasound for cervical lymph node metastasis was 0.667.Through feature selection and model establishment,the best model was selected.The best US-based radiomics model to distinguishing benign and malignant thyroid nodules was Spearman0.95+LASSO+Random Forest,which includes 124 related radiomics features.The AUC of the model in the train set was 0.968,and the AUC in the test set was 0.840.Spearman0.95+LASSO+SVM was the best US-based radiomics model to identify cervical lymph node metastasis of thyroid cancer.A total of 16 related radiomics features were included.The AUC of the model in the train set was 0.787,while the AUC in the test set was 0.733,which was significantly higher than that of conventional ultrasound.Conclusion US-based radiomics provides a novel approach for preoperative thyroid nodules diagnosius and cervical lymph node metastasis prediction.It is expected that expanding the large sample size,conducting multi-center research and prospective research could obtain reliable identification performance,which would be helpful for clinical application.