| Background:Thyroid carcinoma(TC)is a common endocrine malignancy.Its incidence rapidly increased within the past decades.According to the latest data released by international agency for research on cancer(IARC)of the World Health Organization(WHO)in 2020,breast cancer has become the world’s most popular cancer with 2.26 million new cases,while thyroid cancer has 567000 new cases,accounting for 3.1%of all the cancers,ranking ninth.There is also a soaring in China.The report from the National Cancer Center in 2019 showed that the incidence of thyroid cancer in Chinese was 14.6 per 100000 in whole and 22.56 per 100000 in female,significantly higher than the global average level,ranking fourth in female tumors.Ultrasonography(US)is a useful examination for screening of thyroid nodules.The detection rate of high-resolution ultrasonography in random population could reach 19%to 68%.Malignant risk stratification based on US has been widely used in clinical practice.Since 2009,thyroid imaging reporting and data system(TI-RADS)was initially proposed by Horvath proposed,most researchers committed to the development and improvement of TI-RADS,but there is no unified version yet.The version of TI-RADS proposed by Kwak in 2011 and the latest version from American College of Radiology(ACR)in 2017 are widely used in China.However,the diagnostic efficacy of those two versions in Chinese population is unclear.TI-RADS is mainly based on US examination,but in addition to US imaging features,there are also significant differences in clinical characteristics between benign nodules and thyroid cancer.According to previous studies and results of our preliminary work,sex,age,thyroid stimulating hormone(TSH)and preoperative anti thyroglobulin antibody(TgAb)were both correlated with the risk of papillary thyroid carcinoma(PTC).Establishing a forecasting model combining clinical and ultrasonographic features might bring more clinical benefits.Objective:1.Analyzing and comparing the diagnostic efficacy of Kwak TI-RADS and ACR TI-RADS in the evaluation of thyroid nodules based on Chinese population.2.Investigating the clinical risk factors and ultrasonographic features of PTC and constructing a forecasting model by methods of machine learning(ML),providing visual interpretation of the model.Methods:1.The comparative study of clinical application of TI-RADS:Collecting the clinical and ultrasonographic data of patients with thyroid nodules who received thyroidectomy in PLA General Hospital from 2000 to 2015.The thyroid nodules with definite pathological diagnosis were enrolled and graded according to the criteria of Kwak TI-RADS and ACR TI-RADS.Receiver operating characteristic curve(ROC curve)was used to evaluate the diagnostic efficacy.The area under curve(AUC),sensitivity,specificity,positive predictive value,negative predictive value and accuracy between two versions of TI-RADS have been comparatively analyzed.2.Investigating the risk factors of PTC on both clinical and ultrasonographic characteristics,which would be screened as features in forecasting model construction.1484 thyroid nodules with complete clinical and ultrasonographic data were selected as the data set to build forecasting model.The data sets were randomly divided into 70%training set and 30%validation set.AUC,sensitivity,specificity,positive predictive value,negative predictive value and accuracy were used to evaluate the diagnostic effectiveness of the prediction model.The diagnostic efficacy of the predictive model was compared with that of two versions of TI-RADS,and then the visual interpretation of the predictive model was performed.Results:1.A total of 7341 patients with thyroid nodules were enrolled in this study,including 3437 benign and 3904 PTC patients.There were 8830 thyroid nodules,4522 were benign and 4308 were PTC.The results of ROC curve suggested that the AUC of Kwak TI-RADS was 0.83(95%CI:0.82-0.84,P<0.001)and the cut-off point was 4a.When the nodule was graded as 4a or higher levels,the risk of PTC was higher.The sensitivity,specificity,positive predictive value,negative predictive value and accuracy of the cut point were 0.86,0.65,0.70,0.85 and 0.76.The AUC of ACR TI-RADS was 0.79(95%CI:0.78-0.80,P<0.001),and the cut-off point was TR3.When the nodule was graded as TR3 or higher levels,it presented higher risk of malignancy.The sensitivity,specificity,positive predictive value,negative predictive value and accuracy of the cut point were 0.83,0.62,0.56,0.71 and 0.60,respectively.The AUC of Kwak TI-RADS was higher than that of ACR TI-RADS in both whole nodules and subgroups divided by gender,age,maximum diameter(≤1cm and>1cm)and coexistence with Hashimoto’s thyroiditis(HT).The two versions of TI-RADS by Kwak and ACR showed better diagnostic efficiency in patients younger than 55 years old,with nodules larger than 1cm and those without HT(P<0.001).2.The clinical characteristics of 7341 patients with thyroid nodules were analyzed.The results showed that younger than 55 years old,located in single lobe including left lobe,right lobe or isthmus),TSH level and positive TgAb were independent risk factors for PTC.As for ultrasonographic features,nodules with maximum diameter ≤1cm,aspect ratio(AR)>1,irregular shape,unclear boundary,microcalcification,solid composition and hypoechoic were related to a higher risk of malignancy(P<0.001).All the above clinical and ultrasonographic factors of PTC were used as features in forecasting model.Three algorithms of machine learning including Extreme Gradient Boosting(XGBoost),Random Forest(RF),and Logistic Regression(LR)were used to train the predictive model in data set composed of 1484 nodules.ROC was used to evaluate the efficiency of the forecasting model.The AUC of XGBoost,RF and LR were 0.91,0.90 and 0.90,respectively,which were both higher than that of Kwak(0.89)and ACR(0.85)TI-RADS.Conclusion:Kwak TI-RADS and ACR TI-RADS have better clinical application value.On the whole,the diagnostic accuracy of Kwak TI-RADS is higher than that of ACR TI-RADS.The diagnostic efficiency of the two versions is variated in subgroups with different age,nodule size on ultrasound and coexistence with HT.The forecasting model constructed by machine learning showed better diagnostic efficiency compared with Kwak TI-RADS and ACR TI-RADS. |
PDF Full Text Request