Comparison Of Thermal Ablation And Surgery For Benign Thyroid Nodules And The Prediction Of Local Treatment Response Of Thermal Ablation By Ultrasound Radiomics

Purpose:1.This study aimed to evaluate the value of ultrasound(US)-guided thermal ablation(TA)by comparing the efficacy,safety,and patient satisfaction between TA,conventional open thyroidectomy(COT)and endoscopic thyroidectomy(ET)for benign thyroid nodules(BTNs).2.This study aimed to establish a machine learning(ML)-based ultrasound radiomics model for preoperative prediction of the mid-term local treatment response of TA for patients with BTNs.Materials and Methods:1.Patients with BTNs who underwent COT,ET,or TA(radiofrequency ablation,RFA or microwave ablation,MWA)therapy from January 2018 to January 2020 were included.The volume reduction rate(VRR),technique efficacy rate(VRR ≥ 50%),nodule disappearance and regrowth rates were calculated after ablation.Propensity score matching(PSM)was used to balance the pre-operation data of the two comparisons(TA vs.COT and TA vs.ET).The treatment and hospitalisation time,medical cost,complications,postoperative symptoms,and cosmetic scores were recorded and compared.Patient satisfaction was evaluated using a telephone survey.2.A retrospective analysis was performed of patients with BTNs who received TA(RFA or MWA)in three hospitals between January 2018 and July 2021,and all basic information and US images of the patients were collected.Image annotation was performed by two physicians with more than 5 years of US experience using 3D Slicer software,and the radiomics features were extracted from the US images using Py Radiomics.The extracted radiomics features included: first-order features,shape features,grayscale co-occurrence matrix features,grayscale size region matrix features,grayscale run-length matrix features,adjacent grayscale difference matrix features,and grayscale dependency matrix features.The most valuable features were screened by regression fitting to the training set data.The high-dimensional data were reduced by principal component analysis(PCA)and least absolute shrinkage and selection operator(Lasso).Then support vector machine(SVM),logistic regression,decision Tree,K-nearest neighbor,and random forest were applied to distinguish VRR≥ 50% or not at 6 months after ablation.The ML algorithms were trained using a20-fold cross-validation mechanism in the training cohort.Factors affecting VRR <50% were obtained through univariate and multivariate logistic regression analysis and then the clinical model was established.The selected radiomics features were combined with clinical factors to construct a combined model.Receiver operating characteristic(ROC)curve was used to analyze and evaluate the predictive performance of the model.Results:1.A total of 505 patients who received COT(n = 320),ET(n = 56),or TA(n =129)were enrolled in the study.After PSM,two paired groups were formed(118patients in the TA group vs.118 patients in the COT group;43 patients in TA group vs.43 patients in ET group).In the two matched groups,there were no significant differences in the maximum diameter and volume of major nodules,gender,age,symptom and cosmetic scores(all P > 0.05).After a median follow-up for 19 months(range,12–36 months),the mean VRR was 80.7% ± 21.1%(range,-0.0%–100%),the technique efficacy rate,nodule disappearance and recurrence rates were 92.2%(119/129),7.8%(10/129)and 0.8%(1/129),respectively.Patients in the TA group had less treatment time,hospital stay,and medical expenses than those in the COT and ET groups(all P < 0.001).There were no significant differences in the incidence of complications,postoperative symptom scores,cosmetic scores and overall postoperative satisfaction among the groups(all P > 0.05).The incidence of postoperative hypothyroidism was significantly lower in the TA group than that in the COT and ET groups(all P < 0.05).2.A total of 388 patients were eventually enrolled in this study,including 356 patients in institutions 1 and 2 with a total of 372 nodules containing 405 US images for model training and internal testing.The training set and internal test set were randomly assigned 7:3(283:122)based on the number of images.Institution 3included a total of 32 patients with 32 nodules for external testing of the model.At 6months after ablation,78.5%(292/372)of patients in the training and the internal test sets had nodules with VRR ≥ 50% after TA,compared with 59.4%(19/32)in the external test set.1409 features were obtained for each US image,including 16 shapes and radiographic features extracted from the original and derived images.Based on the PCA algorithm,150 relevant radiomics features were selected.Based on the Lasso regression,the optimal λ in the cross-validation was used to select the best radiomics features with non-zero coefficients,and finally the 10 best features were chosen.Among the five algorithms,SVM had the highest area under the curve(AUC)in the internal test set,and the AUC,accuracy,sensitivity,specificity,positive predictive value,and negative predictive value in the validation set were 0.81(95% confidence interval(CI),0.77–0.86),0.82(95% CI,0.80–0.85),0.78(95% CI,0.71–0.86),0.83(95% CI,0.79–0.87),0.53(95% CI,0.48–0.58)and 0.94(95% CI,0.93–0.96),respectively.Solidity was the only factor significantly associated with VRR < 50% in the multivariate logistic regression analysis(odds ratio,2.52;95% CI,1.356–4.683;P < 0.01).A combined model was constructed using the SVM algorithm in ML with10 radiomics features and one clinical factor.In the internal test set,the AUC of the combined model was similar to that of the ML-SVM radiomics model(0.79 vs.0.81,P = 0.899),and both were significantly higher than that of the clinical model(0.79 vs.0.63,P < 0.01;0.81 vs.0.63,P < 0.05).In the external test set,the AUC values of ML-SVM radiomics model and combined model were similar(0.77 vs.0.77,P =0.945),and both were significantly higher than that of the clinical model(0.77 vs.0.54,P < 0.05;0.77 vs.0.54,P < 0.01).Conclusions:1.Compared with COT and ET,TA has comparable efficacy,safety,and patient satisfaction and exhibits better protection of thyroid function for the treatment of BTNs.It can be regarded as a preferential alternative surgical method for patients with BTNs who are unsuitable or unwilling to receive COT and ET treatment.2.The ML-based US radiomics model performs well for predicting the mid-term local treatment response of TA for patients with BTNs,and its performance is significantly better than that of the clinical model,which will provide an objective reference for physicians and patients to choose TA.