Applying Machine Learning Combined With Contrast-enhanced Ultrasound To Differentiate Benign And Malignant Thyroid Nodules Classified As C-TIRADS 4

Objective: The aim of this study is to evaluate the Contrast-enhanced Ultrasound(CEUS)features of thyroid nodules classified as C-TIRADS 4,analyze and summarize the nodular features of different sizes,and to initially explore the diagnostic effect of machine learning(ML)combined with CEUS on thyroid nodules classified as C-TIRADS 4.Methods: Part I: A total of 322 nodules in 322 patients who underwent CEUS before FNA or surgery at our hospital from September 2020 to November 2022.Clinicopathological data included age,gender,nodule size(max-diameter),location,number,Hashimoto background,and BRAFV600 E gene mutation.C-TIRADS ultrasound(US)features included composition,echogenicity,calcification,shape,and margin.The regions of interest were selected to draw the time intensity curve and analyze the quantitative parameters.And then compare with normal thyroid tissue and calculate the difference,absolute value of the difference,relative number of the difference.Univariate analysis was used to evaluate diagnosing malignant thyroid nodules of every parameter and the results of qualitative and quantitative parameters of CEUS were compared among malignant thyroid nodules in different sizes.Univariate significant variables were selected for multivariate analysis to find independent risk factors for malignant nodules classified as C-TIRADS 4.ROC curves were used to evaluate the diagnostic efficacy.Part II: A total of 194 nodules in 194 patients who underwent CEUS before FNA or surgery at our hospital from September 2019 to June 2022.Collecting clinicopathological information,US information,and CEUS information.The following five CEUS key frames were selected “2nd second after the arrival time”,“time to peak” frame,“2nd second after peak” frame,“First-Flash” frame,“Second-Flash” frame.The data set was randomly divided into training cohort and test cohort by 9:1,and the ROI of five CEUS key frames for each patient were manually labeled for feature selection and extraction.The model development used four classifiers: support vector machine(SVM),random forest(RF),logistic regression(LR)and decision tree(DT).Analysis and compare the diagnostic value of each classifier.The best classifier was selected for comparison with radiologist’s reading of the entire CEUS video.Results: Part I: Most of the malignant thyroid nodules classified as C-TIRADS 4category showed hypo-enhancement(146/212,68.9%),heterogeneous enhancement(155/212,73.1%),non-centripetal enhancement(199/212,93.9%),simultaneous wash-in(122/212,57.5%),synchronous wash out(112/212,52.8%)and no peripheral ring enhancement(190/212,89.6%).ΔPI and PIrn of malignant thyroid nodules classified as C-TIRADS 4(-3.21±3.39 d B and-0.16±0.19,respectively)were significantly lower than those of benign nodules(-0.17±2.17 d B,-0.01±0.09,respectively).|ΔPI|,ΔTTP and TTPrn of malignant thyroid nodules classified as C-TIRADS 4(3.62±2.93 d B,0.51±1.26 s and 0.03±0.07,respectively)were significantly higher than those of benign nodules(1.58±1.49 d B,0.09±1.2s and 0.003±0.06,respectively).CEUS features of C-TIRADS 4 were associated with maximum diameter of the nodule,the larger the maximum diameter of malignant nodules was,the smaller the Δ PI value was,which means,the larger of the malignant nodule was,the more it tended to show "iso-enhancement".Aspect ratio greater than 1,blurred margins,low enhancement andΔPI≤-1.68 d B were independent risk factors for malignant thyroid nodules classified as C-TIRADS 4.The sensitivity,specificity,accuracy,PPV,NPV and AUC of CEUS qualitative analysis combined with CEUS quantitative analysis were 85%,82%,84.1%,91.9%,69.5%,and 0.835,respectively.The sensitivity,specificity,accuracy,PPV,NPV and AUC of CEUS analysis combined with 2D US were 94.2%,62%,84.7%,85.6%,81.6%,and 0.781,respectively.Part II: RF classifier achieved the best performance among the four classifiers,with an AUC of 1 in the training cohort and 0.8(0.5-1)in the test cohort,which was higher than the diagnostic efficacy of radiologist’s reading the whole CEUS video(AUC=0.786 vs.AUC=0.8)(P<0.05).In the test cohort,the sensitivity,specificity,accuracy,PPV,NPV,and F1 scores of the RF model were 83%,93%,90%,83%,and 0.93,respectively.Conclusion: The qualitative and quantitative CEUS analysis provides valuable information for the diagnosis of thyroid nodules classified as C-TIRADS 4,and combing with 2D-US can increase the sensitivity,accuracy and NPV of the diagnosis.ML models based on five key-CEUS-frames provide better performance than radiologist’s reading the entire CEUS video,which can help to achieve a more accurate diagnosis of thyroid nodules classified as C-TIRADS 4.