MRI-based Radiomics Model For Predicting Therapeutic Responses Of Uterine Fibroids After Ultrasound-guided High-intensity Focused Ultrasound Ablation

Objective:To investigate the predictive effect of high intensity focused ultrasound(HIFU)ablation of uterine fibroids using a combined model based on magnetic resonance imaging(MRI)radiomics and clinical features.1.Radiomics features were extracted from conventional MR images of uterine fibroids,including T2-weighted imaging(T2WI),T1-weighted imaging(T1WI),contrast-enhanced T1-weighted imaging(CE-T1WI)and apparent diffusion coefficient(ADC)images.The radiomics-clinical fusion model A was established to predict the immediate nonperfused volume rate using machine learning.2.As above,radiomics features extracted from conventional MR images of uterine fibroids were combined with clinical features.the radiomics-clinical fusion model B was established to predict fibroid volume reduction of uterine fibroids after HIFU using machine learning.The mean follow-up time for fibroid volume was 29 ± 9.23 months.3.Based on the radiomics features extracted from conventional MR images of uterine fibroids,the model C for predicting the long-term follow-up treatment outcome of symptomatic single uterine fibroids after HIFU was established by machine learning,and the mean follow-up time was 39 months(from 38 to 48).Method:1.A total of 257 single uterine fibroid patient were included in our study and were randomly divided into training group and test group(80%,20%).Radiomics features were extracted from pre-treatment pelvic T2WI,T1WI,CE-T1WI,and ADC images.Clinical features included age,BMI,uterine volume,uterine fibroid volume,rectus abdominis muscle thickness,subcutaneous fat thickness,abdominal wall thickness,uterine position,fibroid type,fibroid location,T2 signal intensity.Calculated the immediate nonperfused volume rate based on the volume of the non-perfused area immediately after HIFU.Divided the fibroids into two categories according to the nonperfused volume rate of 80%(higher nonperfused volume rate:nonperfused volume rate≥80%;lower nonperfused volume rate:nonperfused volume rate<80%).Firstly,using Synthetic Minority Oversampling Technique(SMOTE)balanced the data,and then use Min-Max method to normalize the data.Pearson correlation coefficient(PCC)was used for feature analysis,and recursive feature elimination(RFE)used for feature selection.Finally,the radiomics-clinical model for nonperfused volume rate prediction were established by naive bayes(NB).Receiver operating characteristic(ROC)curve were listed and the area under the curve(AUC)was calculated to assess predictive efficacy in the training and test datasets.2.The fibroid volume reduction was calculated according to the volume of uterine fibroid in the long-term follow-up after operation.The fibroids were divided into two categories according to the classification standard of 50%about fibroid volume reduction(higher volume response group:fibroid volume reduction≥50%;lower volume reaction group:fibroid volume reduction<50%).The processing method was the same as above.Finally,logistic regression with LASSO constrain(LR-Lasso)classifier was used for modeling.ROC curve was listed and the AUC value calculated to assess predictive efficacy in the training and test datasets.3.One hundred and ten patients with single symptom uterine fibroids were included in the study for predicting the long-term follow-up treatment outcome of HIFU ablation.Fifty-five patients scanned by Achieva 1.5T(Philips Healthineers,Netherlands)scanner were used as the training set.Fifty-seven patients scanned by the Achieva 3.0T(Philips Healthineers,Netherlands)scanner served as the test set.According to the long-term follow-up outcome,fibroids was divided into two categories(non-treatment failure group;treatment failure group).The processing method was the same as above.Finally,Relief was used to select features,and support vector machine(SVM)as the classifier to build the model.ROC curve was listed and the AUC value was calculated to assess predictive efficacy in the training and test datasets.Result1.In the prediction model(A)for nonperfused volume rate of uterine fibroids,the AUC values of the training and the test dataset are 0.773(95%CI,0.707-0.835)and 0.749(95%CI,0.597-0.891),respectively.2.In the prediction model(B)for fibroid volume reduction of uterine fibroids,the AUC values of the training and the test dataset are 0.807(95%CI,0.745-0.868)and 0.817(95%CI,0.686-0.929),respectively.3.The AUC values of the radiomics prediction model(C)based on long-term follow-up treatment outcome are 0.881(95%CI,0.742-0.973)and 0.831(95%CI,0.701-0.931)in the training and the test dataset,respectively.Conclusion:Models based on radiomics combined with clinical characteristics has predictive value on the efficacy of HIFU for uterine fibroids,and can be used to guide radiologists preoperatively evaluate who can achieve good results of HIFU treatment in patients with uterine fibroids.