基于MRI图像的机器学习模型预测子宫肌瘤HIFU消融术后疗效的价值  

The value of machine learning models based on MRI images in predicting the efficacy of uterine fibroids after HIFU ablation surgery

作　　者：何月明[1] 罗锦文[2] 戚莹莹 张婷[1] HE Yueming;LUO Jinwen;QI Yingying;ZHANG Ting(Department of Gynecology,the Fifth Affiliated Hospital of Guangzhou Medical University,Guangdong Guangzhou 510700,China.;Department of Medical Imaging,the Fifth Affiliated Hospital of Guangzhou Medical University,Guangdong Guangzhou 510700,China.)

机构地区：[1]广州医科大学附属第五医院妇科,广东广州510700 [2]广州医科大学附属第五医院医学影像科,广东广州510700

出　　处：《现代肿瘤医学》2024年第6期1093-1100,共8页Journal of Modern Oncology

基　　金：广东省广州市基础研究计划基础与应用基础研究项目(编号:202102080250)。

摘　　要：目的:探讨基于MRI图像的机器学习模型预测子宫肌瘤高强度聚焦超声(HIFU)消融术后疗效的价值。方法:回顾性分析本院接受HIFU消融治疗前、后的108例子宫肌瘤患者MRI图像及临床资料,以术后子宫肌瘤体积消融率(NPVR)65%为界进行分组,进行肌瘤全瘤MRI定量纹理机器学习,使用ITK-SNAP软件勾画每个肌瘤术前MRI的容积感兴趣区(VOI),使用Python对每个VOI提取1834个组学特征,将特征与标注数据拼接,形成数据集,并将数据正则化,通过相关系数Spearman筛选特征,使用随机数字表法随机构建训练集76例(70%)和测试集32例(30%),通过Lasso筛选非零特征构建LR、SVM、KNN、RandomForest、ExtraTrees、XGBoost、LightGBM、MLP机器学习消融疗效预测模型;绘制ROC、DCA曲线分析、混淆矩阵图并计算各模型的AUC、敏感度、特异度、阳性预测值(PPV)、阴性预测值(NPV)、精确率、召回率、F1分数用于评价预测模型诊断性能及临床应用价值。结果:根据消融效果划分消融显著组69例,消融非显著组39例,随机构建训练集76例(70%)和测试集32例(30%),ROC曲线分析中,MRI组学预测模型预测消融显著组在验证集中最高的四个模型分别为LR 0.875(95%CI 0.731~1.000)、MLP 0.875(95%CI 0.714~1.000)、LightGBM 0.853(95%CI 0.689~1.000)、RandomForest 0.848(95%CI 0.676~1.000),且差异有统计学意义(P<0.05)。结论:基于MRI图像纹理构建的机器学习模型在术前子宫肌瘤超声引导下HIFU消融疗效预测效果好,可为子宫肌瘤HIFU术前选择及术后疗效评估提供个性化及量化参考依据。Objective:To investigate the value of MRI image-based machine learning models in predicting the efficacy of HIFU ablation for uterine fibroids.Methods:We retrospectively reviewed the MRI images and clinical data of 108 patients with uterine fibroids before and after HIFU ablation in our institution,grouped them by 65%postoperative volume ablation rate(NPVR),performed whole tumor MRI texture machine learning of fibroids,used ITK-SNAP software to delineate the volumetric region of interest(VOI)on preoperative MRI for each fibroid,extracted 1834 omics features for each VOI using Python,stitched the features with annotated data,to form the dataset and regularize the data,we randomly constructed 76 cases(70%)of the training set and 32 cases(30%)of the test set to screen the features by correlation coefficient Spearman,and LR,SVM,KNN,RandomForest,ExtraTrees,XGBoost,LightGBM,MLP machine to learn the prediction model of ablation efficacy by filtering the nonzero features by Lasso.ROC and DCA curve analysis,confusion matrix plots were drawn and AUC,sensitivity,specificity,positive predictive value(PPV),negative predictive value(NPV),precision rate,recall rate,F1 score of each model were calculated to evaluate the diagnostic performance of the prediction model and its clinical application.Results:After dividing the ablation significant group by ablation effect in 69 patients,ablation non significant group in 39 patients,randomly constructing the training set in 76 patients(70%)and the test set in 32 patients(30%).ROC curve analysis showed that the MRI omics prediction model predicted ablation significant group the top four models in the validation set were LR 0.875(95%CI 0.731~1.000),MLP 0.875(95%CI 0.714~1.000),LightGBM 0.853(95%CI 0.689~1.000),RandomForest 0.848(95%CI 0.676~1.000),and the difference was statistically significant(P<0.05).Conclusion:The machine learning model based on texture construction of MRI images predicts the efficacy of HIFU ablation under ultrasound guidance for uterine fibroids with good performan

关 键 词：磁共振 机器学习 子宫肌瘤 高强度聚焦超声 

分 类 号：R737.33[医药卫生—肿瘤]