集成机器学习模型联合磁共振视觉及DTI直方图预测成人弥漫性胶质瘤IDH表型的研究  

Prediction of adult-type diffuse gliomas IDH phenotype through an ensemble machine learningmodel with integrating of MRI visual and DTI histogram

作　　者：韩鑫 李笑然 卢洁[1,2] HAN Xin;LI Xiaoran;LU Jie(Department of Radiology and Nuclear Medicine,Xuanwu Hospital,Capital Medical University,Beijing 100053,China;Beijing Key Laboratory of Magnetic Resonance Imaging and Brain Informatics,Beijing 100053,China)

机构地区：[1]首都医科大学宣武医院放射与核医学科,北京100053 [2]磁共振成像脑信息学北京市重点实验室,北京100053

出　　处：《磁共振成像》2024年第11期51-59,89,共10页Chinese Journal of Magnetic Resonance Imaging

基　　金：国家重点研发计划项目(编号:2022YFC2406900)。

摘　　要：目的探索基于堆叠泛化整合磁共振视觉和扩散张量成像(diffusion tensor imaging,DTI)直方图的集成学习模型预测成人弥漫性胶质瘤异柠檬酸脱氢酶(isocitrate dehydrogenase,IDH)表型的价值。材料与方法回顾性分析经2021版WHO中枢神经系统肿瘤分类指南确定的106例成人弥漫性胶质瘤常规MRI及DTI影像,对常规MRI进行伦勃朗视觉感受图像(visually accessible Rembrandt images,VASARI)特征评价,并测量DTI图像的各向异性分数、相对各向异性、容积比各向异性和平均扩散率的绝对值、相对值以及对应的直方图特征。采用递归式特征消除(recursive feature elimination,RFE)和Boruta算法在训练集进行特征筛选,并基于堆叠泛化法将VASARI特征、DTI临床参数及DTI直方图的高斯朴素贝叶斯模型(Gaussian Naive Bayes,GNB)与支持向量机(support vector machine,SVM)集成,建立预测成人弥漫性胶质瘤IDH表型的集成式机器学习模型。采用受试者工作特征曲线下面积(area under the curve,AUC)评估各模型性能。结果106例成人弥漫性胶质瘤(50.05±15.17岁,54例男性)包括55例IDH突变型与51例IDH野生型。RFE与Boruta的级联递归降维分别筛选出6个VASARI特征、8个DTI临床参数特征及8个DTI直方图特征建立初级层分类器。基于DTI直方图的初级分类器的AUC最高(0.90/0.87,训练集/测试集),优于DTI临床参数构建的模型(AUC:0.83/0.78,训练集/测试集)和常规MRI视觉特征构建的模型(AUC:0.84/0.66,训练集/测试集)。基于堆叠泛化的集成学习模型预测IDH表型的AUC最高(0.92/0.89,训练集/测试集)。结论基于堆叠泛化的集成机器学习模型融合常规磁共振视觉特征和DTI特征能准确预测成人弥漫性胶质瘤IDH基因表型,辅助临床快速评估成人弥漫性胶质瘤患者的预后。Objective:To investigate the value of ensemble learning model constructed based on MRI visual and diffusion tensor imaging(DTI)histogram for predicting isocitrate dehydrogenase(IDH)phenotypes in adult-type diffuse gliomas.Materials and Methods:A retrospective analysis was conducted on conventional MRI and DTI images of 106 adult diffuse gliomas identified by the 2021 edition of the WHO Classification of Central Nervous System Tumors.Visually accessible Rembrandt images(VASARI)features were evaluated on conventional MRI.The absolute and relative values of fractional anisotropy(FA),relative anisotropy(RA),volume ratio anisotropy(VR),and mean diffusivity(MD)of DTI images were measured,as well as the histogram features.Recursive feature elimination(RFE)and Boruta algorithms were used for feature screening in the training set,and Gaussian Naive Bayes(GNB)models of VASARI features,DTI clinical parameters and DTI histograms were ensembled with a support vector machine(SVM)based on the stacking method.The ensemble machine learning model was then used to predict the IDH phenotype of adult diffuse gliomas.The performance of each model was evaluated by measuring the area under the curve(AUC)of receiver operating characteristic curves.Results:A total of 106 glioma patients(50.05±15.17 years old,54 males)were enrolled in the study,comprising 55 patients with IDH-mutant and 51 patients with IDH wildtype.The cascade recursive dimension reduction of RFE and Boruta,respectively,identified six VASARI features,eight DTI clinical parameters features,and eight DTI histogram features as the primary layer classifier.The model constructed based on histogram features had the highest AUC(0.90/0.87,training set/test set),which was superior to the model constructed from DTI clinical parameters(AUC:0.83/0.78,training dataset/testing dataset)and the model constructed from conventional MRI visual features(AUC:0.84/0.66,training dataset/testing dataset).The ensemble learning model based on stacking generalization achieved the highest AUC for p

关 键 词：胶质瘤 磁共振成像 扩散张量成像 集成机器学习 异柠檬酸脱氢酶 

分 类 号：R445.2[医药卫生—影像医学与核医学] R730.264[医药卫生—诊断学]