The Value Of CT Radiomics In Differential Diagnosis Of Benign And Malignant Thyroid Micronodules

Background In recent years,with the rapid development of imaging equipment and technology(ultrasound,Computed tomography,Magnetic resonance imaging,etc.),the detection rate of thyroid nodules has increased significantly,of which about 5% are malignant.The pathological types of primary thyroid carcinoma can be divided into papillary thyroid carcinoma,medullary carcinoma,follicular cell carcinoma and undifferentiated carcinoma.Most of them are papillary thyroid carcinoma(PTC).Who defines papillary thyroid microcarcinoma as papillary thyroid carcinoma with diameter less than or equal to 1.0 centimeter.Papillary microcarcinoma of thyroid(PTMC)generally grows slowly with good differentiation and good prognosis.However,a small number of thyroid carcinoma has a certain risk of invasion and lymph node metastasis,and the prognosis is relatively poor.How to identify thyroid micronodules,especially for the early diagnosis and treatment of thyroid microcarcinoma and improve the prognosis of patients has important clinical value.objective To investigate the value of CT radiomics in the differential diagnosis of benign and malignant thyroid micronodules.Methods 203 patients with pathologically confirmed thyroid micro-nodules were retrospectively analyzed,including 148 cases of papillary microcarcinoma of thyroid(PTMC)and 55 cases of micro-nodule goiter,MNG),the CT plain scan and dual phase enhanced DICOM images of selected thyroid micro-nodules were processed,the region of interest(ROI)was delineated and 402 texture feature parameters were extracted,including histogram parameters,morphological parameters,gray level cooccurrence matrix parameters(GLCM),run length matrix(RLM),gray level size zoom matrix(GLSZM),Haralick parameter.Twenty,twenty-two,and nineteen feature parameters were obtained by twice feature dimensionality reduction.The following six models were used for evaluation: random forest model,support vector machine(SVM)model,k-nearest neighbor algorithm(KNN)model,decision tree model,bayesian model and logistic model.They were divided into training group and test group: 70% cases in training group and 30% cases in test group were randomly selected.Accuracy,specificity,sensitivity and area under ROC curve(AUC)of different models for differential diagnosis of benign and malignant thyroid micro nodules were obtained.Results Forest was regarded as the best machine learning method among six model for all three phases of CT imaging.The diagnostic accuracy,specificity,sensitivity and area under ROC curve(AUC)of plain scan training group were 99.3%,97.9%,100%and 0.999,and those of test group were 87.9%,81%,91.1% and 0.883,respectively.The diagnostic accuracy,specificity,sensitivity and area under ROC curve(AUC):97.2%,94.4%,100%,0.998 in the arterial phase enhancement training group,77.4%,69.2%,83.3%,0.842 in the test group,100%,100%,100%,1.000 in the delayed phase enhancement training group,82.8%,93.3%,71.4%,0.905 in the test group.Conclusion Radiomics of computed tomography is one of the relatively reliable and quantitative methods to differentiate benign and malignant thyroid micronodules,which can be used as an auxiliary means to identify benign and malignant thyroid micro-nodules;The random forest model of CT imaging has certain value in the differential diagnosis of thyroid microcarcinoma(PTMC)and micro-nodular goiter(MNG).Plain scan and delayed phase enhanced CT imaging have high sensitivity and specificity respectively.