基于影像组学特征的机器学习模型对较低级别胶质瘤ATRX突变状态的预测

Prediction of ATRX mutation status in lower-grade gliomas using a machine learning model based on radiomics features

目的基于多模态影像组学特征构建机器学习模型,探讨其在术前无创性预测异柠檬酸脱氢酶(IDH)突变型较低级别胶质瘤(LrGG)的α-地中海贫血伴智力低下综合征X连锁(ATRX)基因突变状态的能力。方法回顾性分析102例经病理及分子检测确诊为IDH突变型LrGG病人的影像及临床资料,其中ATRX突变型47例,野生型55例。将病人按7∶3的比例随机分为训练集(71例)和测试集(31例)。从增强(CE)-T1WI、表观扩散系数(ADC)图和18F-FDG PET中共提取3318个影像组学特征。根据不同影像来源将影像组学特征分为5个数据集,分别为增强集(CE-T1WI)、ADC集(ADC)、PET集(18F-FDG PET)、MRI集(CE-T1WI+ADC)、联合集(CE-T1WI+ADC+18F-FDG PET)。将4种特征降维方法[线性判别分析(LDA)、主成分分析(PCA)、Wilcoxon基于相关系数选择法、最小绝对收缩和选择算子(LASSO)]以及4种机器学习算法[支持向量机(SVM)、逻辑回归(LR)、K邻近(KNN)、随机森林(RF)]组合,基于联合集特征构建16个预测模型并进行效能评估,选出最优的算法组合。将最优算法应用于增强集、ADC集、PET集、MRI集、联合集中构建模型,绘制受试者操作特征(ROC)曲线并计算曲线下面积(AUC)来评价各模型的预测效能。结果基于联合集影像组学特征构建的16个预测模型中,采用LASSO与RF算法组合构建的模型预测效能最优,其在训练集和测试集中的AUC分别为0.967和0.950。4种特征降维方法中,采用LASSO算法的模型整体预测效能最好;4种机器学习方法中,采用RF算法的模型预测效能最好。该预测模型应用于增强集、ADC集、PET集、MRI集、联合集,结果显示其在联合集中的预测效能最高,在训练集和测试集中的AUC分别为0.967和0.950,其次为MRI集和PET集(AUC值分别为0.931和0.915)。结论基于多模态影像组学特征,采用LASSO与RF算法组合构建的机器学习模型用于预测IDH突变型LrGG的ATRX突变状态的效能较高,是一种无创且简便的方法。

Objective To construct a machine learning model based on multimodal radiomic features and explore its ability to noninvasively predict the mutation status of alpha thalassemia/mental retardation syndrome X-linked(ATRX)gene in isocitrate dehydrogenase(IDH)mutant lower-grade gliomas(LrGG)preoperatively.Methods A retrospective analysis was conducted on the imaging and clinical data of 102 patients pathologically and molecularly confirmed as IDH-mutant LrGG.Of these,47 cases had ATRX mutations,and 55 cases were wild-type.Patients were randomly divided into a training set(71 cases)and a test set(31 cases)in a 7∶3 ratio.A total of 3318 radiomic features were extracted from contrast-enhanced(CE)-T1WI,apparent diffusion coefficient(ADC)maps,and 18F-FDG PET images.The radiomic features were categorized into five datasets based on the imaging source:CE-T1WI dataset,ADC dataset,PET dataset(18F-FDG PET),MRI dataset(CE-T1WI+ADC),and combined dataset(CE-T1WI+ADC+18F-FDG PET).Four feature dimensionality reduction methods[linear discriminant analysis(LDA),principal component analysis(PCA),Wilcoxon-based correlation selection,and least absolute shrinkage and selection operator(LASSO)]and four machine learning algorithms[support vector machine(SVM),logistic regression(LR),K-nearest neighbors(KNN),random forest(RF)]were combined to construct 16 predictive models based on the combined dataset,and their performance was evaluated to determine the optimal algorithm combination.The optimal algorithm was then applied to the CE-T1WI,ADC,PET,MRI,and combined datasets to build models.Receiver operating characteristic(ROC)curves were plotted,and the area under the curve(AUC)was calculated to assess the predictive performance of each model.Results Among the 16 predictive models constructed based on the combined radiomic features,the model combining LASSO with RF had the best predictive performance,with AUCs of 0.967 and 0.950 in the training and test sets,respectively.Among the four feature reduction methods,models using LASSO showed the best overall performance;among the four machine learning algorithms,RF yielded the highest predictive performance.When applied to the CE-T1WI,ADC,PET,MRI,and combined datasets,the model demonstrated the best predictive performance in the combined dataset,with AUCs of 0.967 and 0.950 in the training test and test sets,respectively,followed by the MRI and PET datasets(AUCs of 0.931 and 0.915,respectively).Conclusion The machine learning model combining LASSO and RF algorithms based on multimodal radiomic features has high efficiency in predicting ATRX mutation status in IDH-mutant LrGG.This method is non-invasive and straight forward.