基底节区脑出血血肿周围水肿区的CT影像组学研究

Perihematomal edema in basal ganglia intracerebral hemorrhage by using radiomics approach of CT images

目的探讨CT影像组学技术对基底节区脑出血血肿周围水肿区与正常脑组织的鉴别价值以及其对患者病情评估、预后预测的意义。方法收集西南医科大学附属中医院神经外科自2017年1月至2018年9月收治的120例基底节区脑出血患者的CT图像及临床资料,并将120例患者按3∶1比例随机分组至训练数据集(n=90)和测试数据集(n=30)。应用影像组学分析软件Mazda分别在训练数据集及测试数据集患者的最佳CT图像上的感兴趣区(ROI)内提取出纹理参数,对从训练数据集提取出的纹理参数采用3种降维方法[交互信息法(MI)、Fisher系数法、分类错误概率联合平均相关系数法(POE+ACC)]筛选出特征性纹理参数,再将3种降维方法与4种分析方法[原始数据分析法(RDA)、主成分分析法(PCA)、线性分类分析法(LDA)和非线性分类分析法(NDA)]进行两两组合,建立不同的影像组学标签,以错误率来评价不同标签之间的性能。对从训练数据集获得的特征性纹理参数分别使用R语言软件中随机森林模型、支持向量机模型和神经网络模型进行建模,然后将从测试数据集中提取的纹理参数中导入这些模型中,采用受试者工作特征曲线(ROC)分析这些模型在测试数据集患者中对血肿周围水肿区的预测价值。对所有患者分别依据血肿最大直径、入院时格拉斯哥昏迷评分(GCS)、3个月随访时美国国立卫生研究院卒中量表(NIHSS)评分的中位数分成2组,重复采用Mazda软件分别在2组进行降维处理及建立不同的影像组学标签,以2组错误率之和为总错误率来评价不同标签对患者病情评估、预后预测的意义。结果在90例训练数据集患者的最佳CT图像的ROI内共提取到295个纹理参数,3种降维方法各获得了10个特征性纹理参数。基于POE+ACC法/NDA法这一组合建立的影像组学标签对血肿周围水肿区与正常脑组织鉴别的错误率最低(2.22%)。ROC曲线分析显示,在测试数据集中随机森林模型、支持向量机模型和神经网络模型的曲线下面积分别为0.87(95%CI:0.76~0.97)、0.81(95%CI:0.72~0.93)、0.76(95%CI:0.67~0.89),其中随机森林模型的预测效能最高。基于POE+ACC法/NDA法这一组合建立的影像组学标签对血肿最大直径、入院时GCS评分分析的总错误率最低(26.66%、23.33%),基于Fisher系数法/NDA法这一组合建立的影像组学标签对3个月随访时NIHSS评分分析的总错误率最低(33.33%)。结论采用CT影像组学技术及选择合适的模型进行分析对基底节区脑出血血肿周围水肿区与正常脑组织的鉴别有一定价值,同时对患者病情评估、预后预测有一定意义。

Objective To explore the value of CT images in distinguishing perihematomal edema in basal ganglia intracerebral hemorrhage with normal brain tissue,and its significance in assessing patients'conditions and prognoses.Methods CT images and clinical data of 120 patients with basal ganglia intracerebral hemorrhage admitted to our hospital from January 2017 to September 2018 were collected,and these 120 patients were randomly assigned to group of training data set(n=90)and group of test data set(n=30)at a ratio of 3:1.The texture analysis software Mazda was used to preprocess the CT images and manually sketch the regions of interest(ROIs)to extract the texture parameters in patients from the group of training data set;Mazda software provides texture feature selection methods including mutual information(MI),Fisher coefficients(Fisher),classification error probability combined with average correlation coefficients(POE+ACC),and texture feature analysis including raw data analysis(RDA),principal component analysis(PCA),linear classification analysis(LDA)and nonlinear classification analysis(NDA);texture feature selection methods and texture feature analysis were grouped by pairs to establish different image omics labels;the error rate was used to evaluate the performance of different labels.Random forest model,support vector machine model and neural network model were built for texture parameters in patients from the group of test data set,and texture parameters extracted from patients from group of training data set were imported into these models;receiver operating characteristics curve was used to assess the performance of models.According to the maximum diameter of the hematomas,Glasgow coma scale(GCS)scores at admission,median of National Institute of Health Stroke Scale(NIHSS)scores 3 months after follow up,all patients were divided into two groups;Mazda software was used repeatedly for dimension reduction and establishment of different images omics labels;the sum of error rates from the two groups was taken as total error rate to evaluate the significance of different labels in predicting patients'conditions and prognoses.Results A total of 295 texture parameters were extracted from the ROIs of the best CT images of 90 patients from group of training data set,and 10 characteristic texture parameters were obtained by each of the three dimensionality reduction methods.Among all texture post-processing methods,the lowest error rate was 2.22%for POE+ACC/NDA;AUCs were 0.87(95%CI:0.76-0.97),0.81(95%CI:0.72-0.93)and 0.76(95%CI:0.67-0.89)for random forest model,support vector machine model and neural network model in the test dataset,respectively,which indicated that random forest model had the best forecast performance.The imaging omics labels established based on POE+ACC/NDA had the lowest total error rate for analysis of maximum diameter of hematoma and GCS scores at admission(26.66%,23.33%);the imaging omics labels established based on Fisher's coefficient method and NDA had the lowest total error rate(33.33%)for analysis of NIHSS scores at 3 months of follow up.Conclusion Radiomic method with proper model is of certain value in distinguishing erihematomal edema in basal ganglia intracerebral hemorrhage with normal brain tissue,and also has certain significance in evaluating the patient's conditions and prognoses.