Prediction And Mechanism Study On Prognosis Of High-intensity Focused Ultrasound Ablation Of Uterine Fibroids

High-intensity focused ultrasound(HIFU)has been widely used to treat uterine fibroids,the most common benign gynaecological tumours.As the only non-invasive treatment for uterine fibroids,HIFU ablation is less invasive,safer,has shorter hospitalizations and faster recovery than traditional laparotomy and laparoscopy.However,due to the comprehensive effect of the histopathological characteristics of uterine fibroids,the physical characteristics of HIFU,and the variations in surgical procedures,the prognosis differs among individuals,which further differentiates postoperative follow-up from traditional surgeries.In addition,the removal of space-occupying effect caused by fibroids volume does not occur immediately after HIFU treatment,but only gradually manifests over a period of time.As a result,it requires more frequent follow-up radiology examinations than myomectomy.Contrast-enhanced magnetic resonance imaging(CE-MRI)does not cause radiation exposure and offers the best soft-tissue resolution;therefore,it is an ideal imaging tool for monitoring the prognosis of uterine fibroids after HIFU ablation.However,because of the high cost of MRI examination,long scanning time,and the potential side effects of gadolinium-based contrast agents,it cannot be frequently used for the monitoring of prognosis,which prohibits doctors from being able to understand the changes in postoperative lesions in a timely and accurate manner,thus affecting subsequent clinical decision-making.With the arrival of the medical big data era,machine learning(ML),which has great advantages in analyzing and processing massive data,has gradually become a powerful auxiliary tool for predicting prognosis and making clinical decisions.Through the thinking of big data,the correlation between various independent variables and dependent variables could be found,and rules between imaging features and disease prognosis could be excavated,so as to further establish quantitative prediction models.This will be an ideal way to solve the problem that the follow-up of HIFU ablation of uterine fibroids relies on imaging examination,but cannot be carried out frequently due to the high cost.In this paper,the necrotic tissue and residual fibroids after HIFU ablation were studied in the following three aspects:Objectives(1)The MRI imaging features of HIFU ablation of uterine fibroids were examined and measured by doctors,and different ML models were used to predict the prognosis of necrotic tissue and residual fibroids respectively.The models were evaluated and verified for clinical application.(2)To predict the residual fibroid regrowth within 1 year after HIFU ablation based on T2 weighted imaging(T2WI)radiomics.Establishing the horizontal comparison of clinico-radiological model,radiomics model and radiomics-clinical model,so as to compare the difference between doctors and computers in predicting the prognosis of residual fibroids within 1 year after HIFU ablation of uterine fibroids.(3)Multiple linear regression and case matching were used to analyze the correlation between HIFU dose parameters and necrotic tissue absorption rate in real cases.Corresponding animal experiments were designed to study the mechanism,aiming to provide experimental references for the further search the improvement of clinical symptoms in a shorter time after HIFU ablation of uterine fibroids.Methods(1)The data of patients with uterine fibroids who underwent HIFU ablation at the hospital A and B were retrospectively evaluated.Patients with MRI data at the following time points were included:before treatment,one day post-HIFU,and follow-up.Finally,405 eligible cases from the hospital B were selected as the training set,and 168 eligible cases from the hospital A were used as the testing set.Telephone interview were conducted with all patients for the presence of re-intervention.The following 14 features were obtained by doctors’independent judgment and measurement:age,follow-up time,number of fibroids,number of treated fibroids,FIGO(International Federation of Gynecology and Obstetrics)type,T2WI type,blood supply degree,maximum thickness of residual fibroids,basal distribution of residual fibroids,the non-perfused volume ratio(NPVR),postoperative non-perfused volume(NPV),postoperative residual fibroids volume(RFV),abdominal fat thickness,maximum anteroposterior distance from the surface to the centre of the fibroid.The features that were important for predicting the NPV reduction value and residual fibroid regrowth were selected using the R package‘Boruta’;further,based on the scikit-learn ML library,linear regression(Li R),random forest(RF)and multilayer perceptron(MLP)were used for predicting the NPV reduction value;logistic regression(Lo R),RF,and MLP were used for predicting residual fibroid regrowth.The optimal hyperparameters were selected according to the best coefficient of determination(R~2)in the grid search process by using 10 repeats 10-fold cross-validation,and the average values of the evaluation indexes after 10times of 10-fold cross-validation were taken as the validation results of the models.The R~2,mean absolute error(MAE),mean square error(MSE),median absolute error(Med AE),and explained variance(EV)were used to evaluate the predictive ability of the models for the NPV reduction value.The receiver operating characteristic curve was used to evaluate the predictive ability of the models for residual fibroid regrowth.The optimal models were used to predict the prognosis of 30,60,90,180,270,and 360days after treatment.The prognosis prediction curves were plotted.(2)The data of patients with uterine fibroids who underwent HIFU ablation at the hospital A and B were retrospectively evaluated.Patients with MRI data at the following time points were included:before treatment,one day post-HIFU,and follow-up.Label setting was performed on MRI images one day post-HIFU based on the residual fibroid regrowth within 1 year after treatment.Finally,104 and 324 eligible cases were collected from the two centres respectively,and 243 and 81 cases from the hospital B were used as training cohort and internal validation cohort,and104 cases from the hospital A were taken as external validation cohort.A total of 851 imaging features were extracted from T2WI of all patients.Feature Selection was performed using Least Absolute Shrinkage and Selection Operator(LASSO)in the training cohort.Support Vector machines(SVM)were used to construct clinico-radiological model,radiomics model and radiomics-clinical model respectively.The models were validated by cases from internal and external cohorts.ROC,Area under curve(AUC)and Decision Curve analysis(DCA)were used to evaluate the predictive performance and clinical application value of the models.(3)A total of 209 patients diagnosed with solitary uterine fibroid and received HIFU ablation in the hospital A and B were retrospectively collected.Multiple linear regression was used to analyze the correlation between seven dose-related parameters(including average power,treatment time,irradiation time,treatment intensity,treatment dose,energy efficiency factor(EEF),mass echo change)and the average necrotic absorption rate.The high absorption group and the low absorption group were matched in a ratio of 1:1,the match tolerance was set as 0.06.Matching factors included age,follow-up time,FIGO type,T2WI type,blood supply degree,abdominal fat thickness,maximum anteroposterior distance from the surface to the centre of the fibroid,NPVR,maximum thickness of residual fibroids,basal distribution of residual fibroids,postoperative NPV and RFV.Then,the difference between dose-related parameters and the necrotic absorption rate were further analyzed.Based on the conclusions of clinical data analysis,two kinds of ablation lesion were established in normal liver tissues of experimental rabbits caused by different HIFU doses.The parameters of low dose group were:power 250W,irradiation time 8s,single point single time;the parameters of high dose group were:power 250W,irradiation time8s,single point double times.The prognosis of the ablation lesions was continuously observed by ultrasonography and histopathology immediately,1,3,7,15 and 35 days after ablation,respectively.Results(1)Fourteen MRI features showed predictive importance of NPV reduction value were used to establish the ML models,among which postoperative NPV,postoperative RFV and NPVR ranked in the top three.In the three ML models constructed to predict NPV reduction value,MLP achieved the best prediction performance after 10 repeats 10-fold cross validation,and the average values of R~2,MAE,MSE,Med AE and EV were0.907,7.506,191.089,4.015 and 0.910,respectively.The average R~2 of Li R and RF were 0.809 and 0.879,respectively.In the testing set,Li R achieved the best performance,with R~2,MAE,MSE,Med AE and EV of 0.851,13.375,526.454,7.711 and 0.858,respectively.The R~2 of RF and MLP were 0.828and 0.792,respectively.The maximum thickness of residual fibroids,NPVR and basal distribution of residual fibroids ranked in the top three among the10 features that showed predictive importance of residual fibroid regrowth.Among the three prediction models constructed,RF performed the best in both training and testing set,with an average AUC of 0.904(95%CI,0.869-0.939)after passing 10 repeats 10-fold cross-validation.In the testing set,its AUC was 0.891(95%CI,0.850-0.929),sensitivity was 0.833,specificity was 0.813,and accuracy was 0.821.In the survey of postoperative re-intervention,residual fibroid regrowth in treated uterine fibroids was observed in 174(44.73%)patients,and 74(19.02%)patients received re-intervention,including 24(32.43%)HIFU ablations,41(55.41%)myomectomies,4(5.41%)hysterectomies,3(4.05%)drug therapies,2(2.70%)others.The interval of re-intervention after HIFU ablation was 18(interquartile range:12-36)months.Among the re-intervention patients,77.03%patients were detected residual fibroid regrowth,and the incidence of re-intervention was 32.76%when residual fibroid regrowth occurred;in the absence of regrowth,the incidence was only 7.91%.(2)Among 851 radiomics features extracted based on MR-T2WI,665(78.1%)showed good inter-observer consistency,intraclass correlation coefficient(ICC)>0.8.LASSO was performed to select 17 non-zero coefficient features,including 1 Gray Level Dependence Matrix(GLDM)feature,1 Gray Level Run Length Matrix(GLRLM)feature and 15 wavelet features.The AUC of the prediction models based on T2WI radiomics in the internal and external validation cohorts were 0.830(95%CI,0.733-0.913)and 0.803(95%CI,0.774-0.889)respectively.When radiomics and clinico-radiological features were aggregated,the aggregated model had a higher predictive accuracy than radiomics and clinico-radiological models,with the AUC of 0.922(95%CI,0.857-0.987)and 0.930(95%CI,0.879-0.979)in the internal and external validation cohorts,respectively.DCA indicated that the radiomics-clinical combination model had good clinical application value.(3)In the first section of this part,it was found that EEF was an independent factor influencing the average necrosis absorption rate of uterine fibroids after HIFU ablation(regression coefficient=0.299,P<0.001).The higher EEF was,the higher the average necrotic absorption rate was.The specific relationship was as follows:an increase of 1 in EEF led to an increase of 0.299%in the average necrotic absorption rate.After case matching,EEF was 4.81(2.69-7.05)J/mm~3 in the high absorption group and2.32(1.37-4.55)J/mm~3 in the low absorption group.However,there were no significant differences in average power,treatment time,irradiation time,treatment dose and treatment intensity(P>0.05).In the second section,in the HIFU-induced coagulative necrotic tissue models constructed from normal rabbit liver,the change of the proportion of HIFU damage range in the low dose group(power 250W,irradiation time 8s,single point single time)was higher than that in the high dose group(power250W,irradiation time 8s,single point twice irradiation)within 7 days after irradiation,and decreased more rapidly from the 7th to the 35th day,moreover,EEF in the low dose group was significantly higher than that in the high dose group.Thirty-five days after irradiation,the damaged area in the low dose group was completely replaced by tissue fibers,and were filled with a large number of collagen fiber bundles and surrounded by normal liver tissue.In the high dose group,coagulated necrotic tissue could still be seen in the centre of the damaged area,with scattered light blue calcified particles.Conclusions(1)ML based on MRI features can accurately predict NPV reduction value and residual fibroid regrowth of uterine fibroids after HIFU ablation,and provide a visualization of postoperative changes of lesions by plotting prediction curve of prognosis.This offers a predictive tool for patients to understand the changes of treated fibroids after HIFU ablation,and a new reference method for doctors to manage post-HIFU cases.(2)Radiomics based on MR-T2WI can effectively predict the uterine fibroid regrowth within 1 year after HIFU ablation.It provides an objective,accurate and convenient reference method for clinical decision-making,at the same time,it lays a foundation for the next step to establish a standardized,accurate,reliable and popularizing therapeutic effect evaluation scheme based on big data monitoring.(3)During HIFU ablation of uterine fibroids,the treatment plan should at least leave enough surrounding tissue for tissue repair.A moderate increase on EEF during ablation can promote the absorption of necrotic tissue to a certain extent.The results provide experimental references for further exploring the HIFU ablation scheme of uterine fibroids which can take into account both high NPVR and promoting absorption of necrotic tissue.