【作者】 李静；

【导师】 刘怀军；

【作者基本信息】 河北医科大学 ， 影像医学与核医学（专业学位）， 2020， 博士

【摘要】 脑胶质细胞瘤是最常见的一种中枢神经系统(Central Nervous System,CNS)恶性肿瘤。世界卫生组织(World Health Organization,WHO)将脑胶质细胞瘤分为低级别胶质细胞瘤(Low-grade Gliomas,LGGs,I级或Ⅱ级)和高级别胶质细胞瘤(High-grade Gliomas,HGGs,Ⅲ级或Ⅳ级)。HGGs占所有脑胶质细胞瘤的60～75%,包括胶质母细胞瘤(Glioblastoma,GBM)和间变性胶质细胞瘤(Anaplastic Glioma,AG),GBM年平均发病率为3.19/10万,由于具有较高的侵袭性和增殖潜能,GBM患者的5年生存率低于5%,其平均生存期为诊断后约15个月。LGGs占胶质瘤的30%,其被定义为遗传异质性肿瘤,尽管预后相对良好,但几乎所有LGGs最终都会发生恶性转化。尽管准确的脑胶质细胞瘤分级具有重要的临床价值,但是WHO分级相同的个体间预后却有着较大差异。有充分的证据表明,组织特征相同或相似的脑胶质细胞瘤可以具有不同的分子遗传学背景。2016年WHO通过将分子标记物与组织病理学相结合,更新了脑胶质细胞瘤分类,WHO对CNS肿瘤的分类强调了分子检测的重要性,在最终的组织病理学报告中必须进行分子检测以进行后续临床决策,从而体现分子标记物的重要性。其中,Ki-67标记指数(Ki-67 Labeling Index,Ki-67 Li)、vimentin、CD34及S-100的表达是判断和预测肿瘤细胞生物学行为的重要因子。Ki-67核抗原仅存在于增殖细胞中,是快速评价正常和异常细胞生长率的可靠指标。vimentin作为上皮细胞向间充质细胞分化过程(Epithelial-Mesenchymal Transition,EMT)的生物标志物越来越受到重视,而EMT与肿瘤细胞的运动性和侵袭性密切相关,其表达与星形细胞瘤预后有关。CD34是众所周知的内皮标记物,在生理和病理血管中具有阳性染色,因其具有良好的免疫反应性,也被认为是微血管密度研究的最佳标志物。S-100是一种酸性钙结合蛋白,在脑组织中含量很高,主要位于胶质细胞,其在脑组织中的出现与神经系统的成熟有关。基于肿瘤遗传学水平的免疫组织化学(Immunohistochemistry,IHC)生物标志物的病理分型对于疾病预测、管理和预后的判断准确性和个性化更高,临床价值也更大。目前手术及活检后的病理组织学是IHC标志物检测的金标准,然而其具有侵入性、采样不及时、耗时、取样误差、不同的组织学解释以及临床实践中各IHC数据难以获取齐全等局限,有必要开发一种针对IHC病理生物标志物分型的无创、有效的技术手段,以期更加准确地对不同的脑胶质细胞瘤亚型进行病理分型。磁共振成像(Magnetic Resonance Image,MRI)作为一种非侵入性工具,在脑胶质细胞瘤诊断中起着重要作用。而影像组学作为一种新型的研究技术,与影像数据严密合作,能从影像数据中提取到高通量的组学特征。这些特征在脑胶质细胞瘤的分级研究中已有诸多的文献报道,其中MRI序列的选择较为多样,包括一些常规的序列如T1WI、T2WI和T1对比增强(Contrast-enhanced T1-weighted Images,T1-CE)以及一些先进的技术,如弥散加权成像(Diffusion-weighted Imaging,DWI)和动脉自旋标记(Arterial Spin Labeling,ASL)等,大多研究结果显示,影像组学特征具有很好的脑胶质细胞瘤分级鉴别性能,然而关于液体衰减反转恢复序列(T2-weighted Fluid-attenuated Inversion Recovery,T2 FLAIR)的研究并不多。目前,基于影像组学特征进行脑胶质细胞瘤IHC分型的研究比较少,也没有可靠的定论,同时进行多组IHC数据研究的报道更少。而T2 FLAIR序列能捕捉到一些非强化状态下的特征,它抑制自由水并且可以清楚地显示结合水,可以系统地量化可见的肿瘤表型特征。目前,仅可检测到关于脑胶质细胞瘤Ki-67的影像组学研究,尚未有关于脑胶质细胞瘤S-100,CD34和vimentin这三种IHC指标的影像组学研究。在这项研究中提出了基于影像组学特征和它们的二元逻辑回归模型,来预测Ki-67,S-100,CD34和vimentin的IHC类型,从而为脑胶质细胞瘤提供非侵入性,以及更准确和个性化的疾病管理方案。第一部分T2 FLAIR序列影像组学特征与脑胶质细胞瘤分级联系的研究目的:探讨基于T2 FLAIR序列的高阶影像组学特征与脑胶质细胞瘤分级的联系,并联合临床特征来观察模型性能。方法:回顾性分析2015年3月至2018年6月本院收治的51例经病理证实的脑胶质细胞瘤患者,均进行了T2 FLAIR检查,采用ITK-SNAP(www.itk-snap.org)软件勾画感兴趣容积(Volume of Interest,VOI),应用A.K.(Artificial Intelligence Kit,GE Healthcare)软件提取影像组学特征。采用方差分析+秩和检验(ANOVA+Wilcoxon)、相关性分析(Sperman)以及LASSO算法(Least absolute shrinkage and selection operator,LASSO)、梯度下降算法(Gradient Descent,GBDT)进行特征降维。对二分类分布严重偏倚的样本数据采用SMOTE数据扩增法解决数据分布不平衡。模型评价指标采用ROC曲线下面积(Area Under Curve,AUC)、灵敏度、特异度、准确度。根据决策曲线的标准净收益来确定模型的可靠性。应用影像组学特征对脑胶质细胞瘤的高低级别分级进行分析。结果:1.51例数据中最终获得396个特征,共入组四个特征,包括两个灰度共生矩阵特征(Gray Level Co-occurrence Matrix,GLCM):Harali-ck Correlation＿angle135＿offset7和Inverse Difference Moment＿All Dire-ction＿offset4＿SD;两个灰度游程矩阵特征(Gray Level Run Length Matrix,GLRLM):Low Grey Level Run Emphasis＿All Direction＿offset1＿SD和Short-Run Emphasis＿All Direction＿offset7。GLCM和GLRLM在特征簇中所占的比例为100%。模型拟合优度检验χ~2值为2.797、P=0.946,AUC:0.888,灵敏度:0.781,特异度:0.895,对模型进行校准,平均绝对误差(Mean absolute error)=0.023,绝对误差分位(Quantile of absolute error)=0.049。2.临床特征中年龄呈正态分布(Shapiro-Wilk test,P=0.2353),且在脑胶质细胞瘤高低级别组间的分布差异具有统计学意义(P=0.015),性别在脑胶质细胞瘤高低级别组间差异不具有统计学意义(P=0.489),将年龄与影像组学特征合并能显著提高模型性能,模型拟合优度检验χ~2值为3.477、P=0.901,AUC:0.929,灵敏度:0.938,特异度:0.789,对模型校准,平均绝对误差=0.028,绝对误差分位数=0.061。第二部分T2 FLAIR序列影像组学特征与脑胶质细胞瘤多种免疫组织化学指标联系的研究目的:建立Ki-67、vimentin、S-100、CD34四项IHC阳性/阴性鉴别的逻辑回归模型,验证影像组学特征预测脑胶质细胞瘤分子分型的可行性。方法:收集2015年3月至2018年6月本院收治的51例经病理证实并均进行了T2 FLAIR检查的脑胶质细胞瘤患者的Ki-67、vimentin、S-100、CD34四项IHC数据,应用影像组学方法,筛选出与四组IHC数据的阳性/阴性分类有关的显著预测因子并建立逻辑回归模型。本研究中Ki-67的分类标准根据Ki-67标记指数(Ki-67 Labeling Index,Ki-67 Li)阳性率分为4级:0～5%为0级,6～25%为1级,26～50%为2级,大于50%为3级,将0级和1级视为阴性,将2级和3级视为阳性;还有文献提到第二种Ki-67分类方法是Ki-67 Li<20%划分为阴性,≥20%划分为阳性,同时对第二种方法进行了分析。结果:1.基于Ki-67、S-100、vimentin、CD34四项IHC数据共建立了四组逻辑回归预测模型。年龄和性别在四组间的分布差异均无统计学意义(P均>0.05)。2.Ki-67影像组学模型共筛选出五个特征,包括三个GLCM,一个Haralick和一个GLRLM;S-100影像组学模型共筛选出五个特征,包括一个Histogram,三个GLCM和一个GLRLM;vimentin影像组学模型共筛选出三个特征,包括一个GLCM和两个GLRLM;CD34影像组学模型共筛选出三个特征,包含一个GLCM和两个GLRLM。四组模型中均未纳入形态学特征和灰度区域大小矩阵(Gray Level Size Zone Matrix,GLSZM),然而每个模型都包含了GLCM和GLRLM,且二者的比例在所筛选的特征簇中均在80%及以上(Ki-67:80%,S-100:80%,vimentin:100%,CD34:100%)。四组模型共16个特征之间的相关性均较低(相关系数r<0.5),这表明了每组模型的特征之间以及每组模型的特征簇之间为模型预测性能提供了不同的信息,不同的组学特征提供的信息以及预测效果具有各异性。3.在Ki-67、S-100、vimentin和CD34四种模型中每一个预测模型的Radscore值在阳性/阴性标记间的分布差异具有统计学意义(P均<0.05,Kruskal-Wallis-H检验)。因此,Radscore值在四组模型中均可以作为IHC阳性/阴性预测的显著性因子。4.对四组模型进行Hosmer-Lemeshow拟合优度检验,Ki-67、S-100、vimentin、CD34的χ~2分别为2.975、2.489、6.833和9.214,P值分别为0.936、0.928、0.555、0.325。结果显示,四组预测模型与实际模型差异均无统计学意义(P>0.05),其中S-100模型与实际模型的拟合效果最好,Ki-67、S-100、vimentin三组模型的拟合优度均好于CD34模型。此外,Ki-67、S-100、vimentin、CD34模型的赤池信息准则(Akaike Information Criterion,AIC)分别为:72.509、46.163、45.037、56.654,结果表明,S-100和vimentin模型的拟合优度优于Ki-67和CD34模型。但Ki-67和CD34模型相对不可靠的具体原因有待进一步验证。另外,S-100模型的阳性似然比(Positive Likelihood Ratio,+LR)最高为9.38,阴性似然比(Negative Likelihood Ratio,-LR)最小为0.12,表明模型在预测S-100蛋白的正负表达时正确判断的概率远大于错误判断的概率。较高的阳性预测值92.6和阴性预测值82.4也表明S-100模型预测的准确性较高。vimentin模型也具有较高的预测性能。然而,在Ki-67和CD34模型中,综合指标的预测效果相对较差。5.T2 FLAIR序列影像组学特征区分Ki-67,S-100,vimentin和CD34模型的AUC分别为0.713、0.923、0.854和0.745;灵敏度分别为0.692、0.893、0.875和0.556;特异度分别为:0.667、0.905、0.722和0.875,准确度分别为0.660、0.898、0.738和0.667。根据决策曲线Ki-67,S-100和vimentin模型具有参考价值。6.当将Ki-67 Li<20%划分为阴性,≥20%划分为阳性时,Ki-67模型进行Hosmer-Lemeshow拟合优度检验,χ~2为7.779,P值为0.455;AIC值为45.124;AUC为0.916;灵敏度为0.849,特异度为0.882,准确度为0.820。第三部分T1-CE序列影像组学特征在Ⅲ级和Ⅳ级脑胶质细胞瘤的实体瘤和水肿区域的比较研究目的:对23例T1-CE序列的实体瘤和水肿区域的Ⅲ级和Ⅳ级脑胶质细胞瘤的影像组学特征进行比较分析。方法:在上述收集的51例患者中有23例进行了T1-CE检查,采用ITK-SNAP软件对Ⅲ级和Ⅳ级脑胶质细胞瘤的T1-CE图像中增强的实体瘤及水肿区域进行感兴趣勾画,应用影像组学方法对Ⅲ级和Ⅳ级脑胶质细胞瘤的实体瘤及水肿区域有无区别进行分析。结果:1.经过特征筛选后,实体瘤中,可区分Ⅲ级和Ⅳ级脑胶质细胞瘤的特征簇包含一个直方图特征(Histogram):Mean Deviation,一个形态学特征:Compactness1,三个GLCM:Cluster Prominence＿angle90＿offset7、GLCME-ntropy＿All Direction＿offset1、GLCMEntropy＿angle0＿offset4和一个GLR-LM:Long Run Emphasis＿angle90＿offset1总共6个特征,其中GLCM和GLRLM在特征簇中所占的比例为66.7%。2.水肿区域中,Ⅲ级和Ⅳ级筛选的特征簇包括一个Histogram:Energy,2个GLCM:Cluster Prominence＿All Direction＿offset1＿SD、GLCME-ntropy＿angle135＿offset1和2个GLRLM:Long-Run Emphasis＿angle135＿offset7、Short-Run High Grey-Level Emphasis＿All Direction＿offset7＿SD,同样,GLCM和GLRLM在特征簇中所占的比例为80%。3.Ⅲ级和Ⅳ级脑胶质细胞瘤实体瘤和水肿区域的直方图特征分布的均匀性、峰度、偏度等均有很大差异。结论:1.本研究结果表明,影像组学特征能很好地对脑胶质细胞瘤高低级别分级做出鉴别,将年龄与影像组学特征合并能显著提高模型性能。2.基于T2 FLAIR的影像组学特征可以预测IHC指标Ki-67、S-100、vimentin和CD34的表达。而CD34模型的正、负判别能力较弱,这可能与本研究所包含的数据分布有关。影像组学模型有望成为脑胶质细胞瘤计算机-智能化的、无创的、精确的、个性化的疾病管理方法。3.Ⅲ级和Ⅳ级脑胶质细胞瘤实体瘤和水肿区域的直方图特征分布的均匀性、峰度、偏度等存在很大差异;实体瘤及水肿区域中,可区分Ⅲ级和Ⅳ级脑胶质细胞瘤的特征簇中GLCM和GLRLM所占比例均较高。4.将Ki-67 Li以20%为界划分阴性/阳性时,Ki-67模型的预测效果高于以25%为界划分阴性/阳性时。更多还原

【Abstract】 Glioma is the most common malignant tumor of the central nervous system(CNS).According to the World Health Organization(WHO)criteria,it can be divided into low-grade gliomas(LGGs,grade I or II)and high-grade gliomas(HGGs,grade Ⅲ or Ⅳ).HGGs which include glioblastoma(GBM)and anaplastic glioma(AG)account for 60～75%of all gliomas.The average annual incidence of GBM is 3.19/100,000.Due to the high invasive and proliferative potential,the 5-year survival rate of glioblastoma patients is less than 5%,and a median survival duration is approximately 15 months after diagnosis.30%of gliomas are LGGs,which is defined as hereditary heterogeneous tumors.Although the prognosis is relatively good,almost all LGGs eventually undergo malignant transformation.Accurate grading of glioma is meaningful for clinical value,but significantly different prognosis exists among individuals who were classified as the same WHO grade.It has been fully studied that gliomas with the same or similar histological characteristics may carry different molecular genetic information.In 2016,WHO updated the classification of glioma by combining molecular markers with histopathology.WHO emphasized the importance of molecular detection in the classification of CNS tumors.In the final histopathological report,molecular detection must be performed to make follow-up clinical decisions and the importance of molecular markers can be demonstrated.Ki-67 labelling index(Ki-67 Li),vimentin,CD34 and S-100 are vital biological behavior biomarkers.Ki-67 nuclear antigen only expresses in proliferating cells,which makes it a reliable avenue to rapidly evaluate the growth fraction of normal and abnormal cells.Vimentin has gained more attention as an emblematical biomarker for Epithelial-mesenchymal transition(EMT),a process of epithelial cells to mesenchymal cells trans-differentiation,and EMT has a close association with tumor motility and invasiveness.CD34is well known as an endothelial marker which presents positive staining in physiologic and pathologic vessels.CD34 is also considered to be an optimum marker of microvascular density studies because of its good immunoreactivity.S-100 is an acidic calcium binding protein found primarily in glial cells in brain and correlates with maturation of the nervous system.Pathological typing of immunohistochemical biomarkers based on tumor genetics is more accurate and individualized for disease diagnosis,management and prognosis prediction which has significant clinical value.At present,the pathological histology of glioma after surgical resection or biopsy is the golden standard for gliomas grading and immunohistochemical typing.However,it has some inadequacies such as invasiveness,untimely sampling,time-consuming,sampling errors,different histological interpretations and limitations of obtaining complete IHC data in clinical practice.Therefore,it is necessary to find an effective and non-invasive approach to classify different glioma immunohistochemical subtypes.Magnetic resonance image(MRI),a non-invasive tool,plays an important role in the diagnosis of glioma.Radiomics,a new research technology,works closely with image data and can extract high-throughput features from medical images.Radiomics features have been used to predict grades of gliomas and show good performance.The sequences of MRI in the previous studies include:conventional sequences such as T1WI,T2WI and Contrast-enhanced T1-weighted images(T1-CE),advanced MRI techniques such as diffusion-weighted imaging(DWI)and arterial spin labeling(ASL)etc.However,T2-weighted fluid-attenuated inversion recovery(T2 FLAIR)was studied rarely,let alone the glioma immunohistochemical typing research based on this.T2 FLAIR can capture some features in the non-enhanced state,and the visible tumor phenotypic characteristics can be systematically quantified.It limits free water and can clearly show bound water.At present,only Ki-67radiomics studies can be detected,but S-100,CD34 and vimentin have not been discussed yet.In this study,we proposed radiomics features and the binary logistic regression model based on them to predict the immunohistochemical typing of Ki-67,S-100,CD34 and vimentin,so as to provide a non-invasive,more accurate and personalized disease management for glioma patients.Part One Study on the relationship between the radiomics features and glioma grades based on T2 FLAIR.Objective:To explore the relationship between the high-order radiomics features and glioma grades based on T2 FLAIR,and to observe the performance of the model combined with clinical features.Methods:51 pathologically confirmed gliomas patients admitted to our hospital from March 2015 to June 2018 were retrospectively analyzed,and all T2 FLAIR imaging were collected.The volumes of interest(VOIs)were manually sketched and the radiomics features were extracted.Feature reduction was performed by ANOVA+Mann-Whiney,spearman correlation analysis,least absolute shrinkage and selection operator(LASSO)and Gradient descent algorithm(GBDT).SMOTE technique was used to solve the data bias between two groups.Comprehensive binary logistic regression models were established.Area under the ROC curves(AUC),sensitivity,specificity and accuracy were used to evaluate the predict performance of models.Models reliability were decided according to the standard net benefit of the decision curves.We analyzed the correlation of radiomics features and glioma grades.Results:1. A total of 396 features were obtained from 51 cases of data.Four features were included in the identification model finally.They were two GLCM features(Haralick Correlation＿angle135＿offset7 and Inverse Difference Moment＿All Direction＿offset4＿SD)and two GLRLM features(Low Grey Level Run Emphasis＿All Direction＿offset1＿SD and Short Run Emphasis＿Direction＿offset7).The proportion of GLCM and GLRLM in feature clusters was 100%.The chi-square value of fit-goodness in this model was 2.797,P=0.946,AUC:0.888,sensitivity:0.781,specificity:0.895.The calibration parameters were mean absolute error=0.023,quantile of absolute error=0.049.2.In addition,we found that the age was normal distribution(Shapiro-Wilk test,P=0.2353),and were statistically different between the high and low grades of glioma(P=0.015),while there was no statistical difference in gender between the two groups(P=0.489).Combining age and radiomics features could significantly improve the model performance.The Chi-square value of fit-goodness of this model was 3.477,P=0.901,AUC:0.929,sensitivity:0.938,specificity:0.789.Additionally,the calibration parameters were mean absolute error=0.028,quantile of absolute error=0.061.Part Two Study on the relationship between T2 FLAIR sequence parameters of glioma.Objective:To establish a logistic regression model for positive/negative identification of IHC by Ki-67,vimentin,S-100 and CD34,to verify the feasibility of predicting molecular typing of glioma by imaging characteristics.Methods:51 pathologically confirmed gliomas patients admitted to our hospital from March 2015 to June 2018 took T2 FLAIR imaging examination,and Ki-67,vimentin,S-100 and CD34 immunohistochemical data were collected.Four groups of logistic regression models for pathological biomarkers positive/negative differential diagnosis were established to verify the feasibility of predicting molecular typing of glioma by radiomics features.In our study,Ki-67 Li was divided into 4 levels according to the positive rate:0～5%was level 0,6～25%was level 1,26～50%was level 2,and more than50%was level 3.In this study,the Ki-67 cohort was divided into label 0 group(level 0 and level 1,negative expression)and label 1 group(level 2 and level 3,positive expression).It was also mentioned in the literature that the second Ki-67 classification method was that Ki-67 Li<20%was divided into label 0and≥20%was divided into label 1.We also analyzed the second method.Results:1. Based on the immunohistochemical data of Ki-67,S-100,vimentin and CD34,logistic regression prediction of four models were established.There was no statistical difference in the distribution of age and gender among the four models(P>0.05).2. We proposed four comprehensive models in revealing immunohi-stochemical typing of Ki-67,S-100,vimentin and CD34.Ki-67 model was composed of five features—three GLCM,one Haralick and one GLRLM;S-100 model included five features—one Histogram,three GLCM and one GLRLM;And vimentin radiomics model enrolled three features—one GLCM and two GLRLM;CD34 model was composed of three features—one GLCM and two GLRLM.Form Factor features were not included in the four models,GLCM and GLRLM were included in each model,and the ratio of them was relatively high in the corresponding feature clusters(Ki-67:60%;S-100:80%;vimentin:100%;CD34:100%).The low correlation coefficients between the16 features indicated little redundancy among every feature cluster.It also suggested that the information and predictive effects provided by single radiomics feature were independent and unique.3. Radscore of each model were significantly different in two labels(all P<0.05,Kruskal-Wallis H test).Therefore,the values of Radscore can be used as a significant factor in immunohistochemical classification in all four models.4. Hosmer-Lemeshow tests were conducted for fit-goodness testing of four models.Theχ~2values of Ki-67,S-100,vimentin and CD34 were 2.975,2.489,6.833 and 9.214,respectively.P values were 0.936,0.928,0.555 and0.325,respectively.Results showed that there was no significant difference between the four classification models and the corresponding actual models.Among them,the S-100 model and the actual model had the best fit-goodness.In addition,the Akaike information criterion(AIC)of Ki-67,S-100,vimentin and CD34 models were 72.509,46.163,45.037 and 56.654,respectively.The results showed that the fit-goodness of S-100 and vimentin models were better than that of Ki-67 and CD34 models.However,the specific reasons for the relative unreliability of Ki-67 and CD34 models need to be further verified.In addition,the S-100 model had the highest positive likelihood(9.38)ratio and the smallest negative likelihood ratio(0.12),indicating that the probability of the correct judgement using model when predicting the positive and negative expression of S-100 protein was much greater than the wrong judgment.Both higher positive predictive values(92.6)and negative predictive values(82.4)also indicate higher accuracy for S-100 model predictions.The high predictive performance was followed by vimentin model.However,in the Ki-67 and CD34 models,the predicting performance were relatively poor in terms of comprehensive indicators.5. Four clusters of significant features were screened out and four predicting models were constructed.AUC of Ki-67,S-100,vimentin and CD34 models were 0.713,0.923,0.854 and 0.745,respectively.The sensitivities were 0.692,0.893,0.875 and 0.556,respectively.The specificities were:0.667,0.905,0.722,and 0.875,with accuracy of 0.660,0.898,0.738,and 0.667,respectively.According to the decision curves,the Ki-67,S-100and vimentin models had reference values.6. When Ki-67 Li<20%was classified as negative and≥20%was classified as positive,the Ki-67 model was tested by the Hosmer-Lemeshow goodness-of-fit test,χ~2was 7.779 and P value was 0.455,AIC value was45.124,AUC was 0.916,sensitivity was 0.849,specificity was 0.882,and accuracy was 0.820.Part Three Radiomics features of solid tumor and edema in grade Ⅲ and Ⅳ glioma through T1-CE imagingObjective:To compare and analyze radiomics features of solid tumor and edema in grade Ⅲ and Ⅳ glioma through T1-CE imaging.Methods:Among the 51 patients collected above,23 cases were examined for T1-CE.ITK-SNAP software was used to sketch the enhanced solid tumor and edema area in T1-CE imaging of grade Ⅲ and grade Ⅳ glioma,and we compared the radiomics features of solid tumor and edema in grade Ⅲ and Ⅳ glioma through T1-CE imaging after feature screening.Results:1.Feature clusters were extracted from solid tumor for distinguishing grade Ⅲ and Ⅳ glioma,containing one histogram feature,one morphological feature(Compactness 1),three GLCM(Cluster Prominence＿angle90＿offset7,GLCMEntropy＿All Direction＿offset1,GLCMEropnty＿angle0＿offset4)and one GLRLM Long Run Emphasis＿angle90＿offset1).Six features were screened out,in which GLCM and GLRLM account for 66.7%of the feature clusters.2.In addition,feature clusters extracted from peritumoral edema includedonehistogramenergy,twoGLCM(Cluster Prominence＿All Direction＿offset1＿SD,GLCMEntropy＿angle135＿offset1)and two GLRLM(Long Run Emphasis＿angle135＿offset7,Short Run High Grey Level Emphasis＿All Direction＿offset7＿SD),similarly,GLCM and GLRLM account for 80%.3.It can be seen from the histogram of characteristic distribution of solid tumors and edema that the uniformity,kurtosis and skewness of feature distribution are quite different between grade Ⅲ and Ⅳ glioma.Conclusion:Overall,our results show that radiomics features are significantly correlated with glioma grade,combining age and radiomics features can significantly improve the model performance.In this study,radiomics features and the binary logistic regression model can predict the immunohistochemical typing of Ki-67,S-100,CD34 and vimentin,so as to provide a non-invasive,more accurate and personalized disease management for glioma patients,while the CD34 model was weaker for positive and negative discrimination,which may be related to the data distribution included in this study.Moreover,it can be seen from the graph that the uniformity,kurtosis and skewness of feature distribution are quite different between grade Ⅲ and Ⅳ glioma.In solid tumors and edema areas,the proportion of GLCM and GLRLM in the characteristic clusters that can distinguish grade Ⅲ and grade Ⅳ glioma is higher.And we have found that when Ki-67 Li is classified as negative/positive by 20%,the prediction effect of Ki-67 model is higher than that by 25%.更多还原