MRI影像组学机器学习模型对肝泡型包虫病边缘带微血管侵犯的预测价值

Evaluating MRI omics-based machine learning for preoperative prediction of microvascular invasion in hepatic alveolar echinococcosis margins

目的探讨MRI影像组学机器学习模型对肝泡型包虫病(HAE)边缘带微血管侵犯的预测价值。方法采用回顾性分析法选取110例经术后病理确诊的HAE患者,收集患者的MR影像资料及病理切片资料,提取病灶MR影像组学特征。将患者按照8∶2比例分为训练集(88例),测试集(22例),建立极限梯度增强树(XGBoost)、随机森林(RF)、Logistic回归(LR)、支持向量机(SVM)和经典决策树(cDT)预测HAE边缘带微血管侵犯的机器学习模型,进行模型验证,并绘制ROC曲线分析不同模型对HAE边缘带微血管侵犯的预测效能。结果病理结果显示110例患者中有25例存在边缘带微血管侵犯,无边缘带微血管侵犯85例。有边缘带微血管侵犯与无边缘带微血管侵犯HAE患者的性别、年龄、病灶部位及病灶大小对比,差异均无统计学意义(P>0.05)。训练集中有20例存在边缘带微血管侵犯,测试集中有5例存在边缘带微血管侵犯。影像组学特征共提取出759个,经方差阈值法、单变量选择法最终筛选出最优影像组学特征6个。训练集中,XGBoost、RF对边缘微血管侵犯诊断的AUC值最高(0.95和0.96);测试集中,XGBoost、RF对边缘微血管侵犯诊断的AUC值(0.88和0.84)也高于其他3组模型。XGBoost模型与RF模型在训练集及测试集中的AUC值比较,差异均无统计学意义(P>0.05)。结论MRI影像组学的XGBoost模型和RF模型对HAE边缘带微血管侵犯的预测均有较高的效能。

Objective To investigate the utility of an MRI imaging omics-based machine learning model for the preoperative prediction of microvascular invasion in the peripheral zone of hepatic alveolar echinococcosis(HAE).Methods A retrospective analysis was conducted on 200 patients diagnosed with HAE,confirmed via postoperative pathology.MR imaging and histopathological data were collected,with histological characteristics of the lesion being extracted from the MR images.The cohort was divided into a training set of 88 patients and test set of 22 patients,following a 8∶2 ratio.Machine learning models,including XGBoost,Random forest(RF),Logistic Regression,Support Vector Machine,and Classical Decision Tree,were developed to predict microvascular invasion in the peripheral zone of HAE.Model validation was performed,and Receiver operating characteristic(ROC)curves were generated to evaluate the predictive performance of the various models on microvascular invasion in the marginal zone of HAE.Results Pathological analysis revealed that out of 200 patients,75 exhibited microvascular invasion in the marginal zone,whereas 125 did not.Statistical analysis indicated no significant differences in gender,age,lesion location,and lesion size between patients with and without marginal microvascular invasion(P>0.05).The training set included 60 cases of marginal zone microvascular invasion,with the remaining 15 cases in the test set.From the collected data,1380 imageomics features were intitially extracted,out of which 406 features were retained using the variance threshold method.Subsequently,six optimal imageomics features were identified through univariate selection,comprising 1 first-order statistical feature and 5 higher-order statistical features.The process of feature selection was refined to ultimately select 6 optimal imaging omics features employing bothe the square difference threshold and univariate selection methods.In the evaluation of diagnostic performance,the XGBoost and RF models demonstrated the highest AUC values in the training set for the diagnosing marginal microvascular invasion,registering AUC values of 0.95 and 0.96,respectively.In the test set,the AUC values for XGBoost and RF were also superior to those of the other model groups,at 0.88 and 0.84 respectively,indicating a higher diagnostic accuracy.Statistical comparison between the AUC values of the XGBoost and RF models in both training and testing sets revealed no significant difference(P>0.05).Conclusion In the preoperative forecasting of microvascular invasion within the marginal zone of HAE,models based on MRI imaging omics,specifically XGBoost and RF,demonstrated superior predictive capabilities.