Construction And Validation Of Two-level CT-based Radiomics Models Used For Thyroid Cancer Screening

Objective: With the increasing incidence of thyroid cancer,the most clinically significant emphasis on the early detection of nodules and accurate determination of malignant ones.Our research aimed to construct and validate unenhanced CT-based radiomics models used for early thyroid cancer screening.Methods: From August 2018 to January 2019,we retrospectively collected CT images of 1356 patients who underwent CT examination as our samples,which contain 343 benign,512 malignant,and 501 healthy control cases respectively.We randomly divided patients into a training set and validation set at a ratio of 7:3.Then,for each patient,396 radiomic features were extracted from the CT-plain images.In the training group,minimum redundancy maximum correlation method(m RMR)and LASSO regression with ten-folds cross validation were adopted as the feature-selection methods to identify suitable features and train the classifiers,and the validation group was independently used to evaluate the established classifiers’ predictive performance.Finally,two-level models were built to detect thyroid nodule and to predict benign or malignant nodule respectively.Firstly,a SVM model was built to the thyroid nodule recognition.Secondly,seven models that incloud RF,c-tree,qbm,qlm,knn,nnet,rf and svm models,evaluated in terms of their stability with relative standard deviations(RSD),were built to predict benign or malignant of the nodule.The diagnostic presentation of every model was evaluated with the index of area under the receiver operating characteristic curve(AUC)Results: In the intra-observer agreement,the average value of ICC was 0.97(range,0.939 to 0.990,p < 0.001)In the inter-observer agreement,the average value of ICC was 0.95(range,0.885 to 0.981,p < 0.001).At the first level,the Support vector machine(SVM)model were built to the thyroid nodule recognition and it presented good discrimination in both the training group(AUC,1.00;95%CI,1.00–1.00)and validation group(AUC,1.00;95% CI,0.99–1.00).At the second level,seven different models were built to predict benign or malignant nodules.Random forest(RF)showed the highest robustness,and its diagnostic performance was also feasible in the validation set(AUC 0.82,95% CI,0.77-0.88).Conclusion: The result of ICC shows that all of the features extracted from the ROI were in good consistency,reproducibility,and stability.Radiomics-based machine learning model could potentially serve as a novel and a promise method to assist radiologists about recognizing the nodules in CT image and to improve the accuracy and efficiency about differentiating malignant thyroid nodules from benign one.