【作者】 王佳；

【导师】 胡春洪；

【作者基本信息】 苏州大学 ， 影像医学与核医学（专业学位）， 2020， 硕士

【摘要】 目的:评估基于扩散张量成像(DTI)的放射组学特征对星形细胞瘤的异柠檬酸脱氢酶1(IDH1)突变状态的预测准确性,并与常规MRI序列和临床特征相比。方法:回顾性纳入了 69例经病理证实为星形细胞瘤的患者,其中男37例,女32例,年龄7～70岁,平均49.39±15.06岁,突变型22例,野生型47例。依据时间将数据集分为训练组和验证组。使用3D-Slicer软件进行图像处理和特征提取,分别从T1WI,T2WI和DTI中提取了共372个放射组学特征。对特征值作Z-Score归一化,单变量分析比较组间差异,对差异具有统计学意义的特征使用ROC曲线分析特征的鉴别价值。使用基于随机森林模型的递归特征消除法及交叉验证筛选最优特征子集,构建放射组学标签。以不同的变量组合构建用于预测IDH1基因状态的模型:(1)临床模型(基于临床和形态学特征);(2)常规放射组学模型(T1WI、T2WI);(3)DTI放射组学模型(ADC、FA);(4)临床-放射组学模型(基于临床模型和放射组学标签)。通过受试者工作特征曲线、临床决策曲线和影响曲线比较模型的预测性能。结果:单变量分析表明372个放射组学特征中一阶统计学特征和灰度共生矩阵(GLCM)特征提供了最有价值的特征参数(AUC:0.696～0.811)。放射组学分析表明ADC衍生的最优特征子集对IDH1基因状态的预测准确性高于T1WI、T2WI和FA(准确性分别为 0.786±0.121,0.713±0.131,0.674±0.242 和 0.613±0.224)。临床特征中年龄[OR(95%CI):0.028(0.002-0.336)]、强化界限[OR(95%CI):0.058(0.005-0.664)]是 IDH1 基因突变的独立危险因子。由基于ADC的放射组学标签和临床特征组成的临床-放射组学模型的预测性能优于单-的临床模型和放射组学标签。结论:直方图和纹理特征对呈形细胞瘤IDH1基因状态有一定的鉴别价值;基于DTI的临床-放射组学模型有助于改善星形细胞瘤患者IDH1基因状态的预测准确性。更多还原

【Abstract】 Purpose:The predictive accuracy of isocitrate dehydrogenase-1(IDH1)mutation status in astrocytoma using radiometric analysis based on diffusion tensor imaging(DTI)was evaluated and was compared with conventional MRI sequences and clinical features.Methods:A number of 69 patients with pathologically confirmed astrocytoma were retrospectively included,of them,there were 37 males and 32 females,aged 7 to 70 years,with an average age of(49.39±15.06)years,22 cases of mutant type and 47 cases of wild type.According to time,the training group and validation group were divided from the data.3D-Slicer software was used for image process and feature extraction.From T1WI,T2WI and DTI,a number of 372 radiomics features were extracted.Comparing the differences between the groups by univariate analysis.The features with statistical significance for the differences were analyzed using the ROC curve to identify the value of discrimination.A recursive feature elimination method based on a random forest model and cross-validation method was used to screen the most feature subsets for constructing radiomics signature.Prediction models for IDH1 gene status was constructed using the following combination of variables:(1)clinical model(based on clinical and morphological features);(2)conventional radiomics model(T1WI,T2WI);(3)DTI-based radiomics model(ADC,FA);(4)Clinical-radiomics model(based on clinical and radiomics)).The models are verified through cross-validation.The receiver operating characteristic curve,clinical decision curve,and impact curve were used to evaluate the predictive performance of the models.Results:Of the total 372 radiomics features,the first-order statistical features and the gray level co-occurrence matrix(GLCM)features provided the most valuable features.The accuracy of the IDH1 genotyping based on ADC-derived optimal feature subset was higher than that of T1WI,T2WI,and FA(accuracy were 0.786±0.121,0.713±0.131,0.674±0.242,and 0.613±0.224 respectively).Among the clinical features,age[OR(95%CI):0.028(0.002-0.336)]and enhancement boundary[OR(95%CI):0.058(0.005-0.664)]were independent risk factors for IDH1 gene mutation.The combined model composed of ADC-based radiomic signature and clinical features performed significantly better than a single clinical model and radiomics model.Conclusions:Histogram and texture features have certain discriminative value for astrocytoma IDH1 gene status.DTI-based radiomics can help improve the predictive accuracy of IDH1 gene status in astrocytoma patients.更多还原