摘要： 目的 探讨扩散峰度成像（DKI）直方图联合红细胞生成素生成的肝细胞受体 A2（EphA2）在胶质瘤分级诊断中的价值。方法 纳入2015 年12 月至2019 年12 月在东部战区总医院行神经外科手术切除并经病理证实的183 例弥漫性胶质瘤患者（包括低级别胶质瘤63 例、高级别胶质瘤120 例），均行常规 MRI 和 DKI 检查［包括部分各向异性（FA）、平均扩散率（MD）、峰度各向异性（KFA）、平均峰度（MK）、平均峰度张量（MKT）］，获取DKI 直方图参数（包括平均值、方差、中位数、25% 分位数、75% 分位数、偏度、峰度），并行 EphA2 免疫组化染色。单因素和多因素 Logistic 回归分析筛查胶质瘤分级预测因素，并基于该因素构建 DKI 直方图以及 DKI 直方图联合 EphA2 评分的胶质瘤分级诊断预测模型，绘制受试者工作特征（ROC）曲线评估其诊断效能，Spearman 秩相关分析探讨 DKI 直方图各项参数与 EphA2 评分的相关性。结果 高级别胶质瘤 FA 值方差（t = -2.050，P = 0.042）和 75% 分位数（t = -2.130，P = 0.035），MD 值方差（t = -6.052，P = 0.000）和偏度（Z = -3.326，P = 0.001），MK 值平均值（t = -3.094，P =0.002）、方差（t = -2.140，P = 0.027）、中位数（Z = -3.444，P = 0.001）、25% 分位数（t = -3.022，P = 0.003）和 75% 分位数（t = -3.438，P = 0.001），MKT 值平均值（t = -3.096，P = 0.002）、方差（t = -2.218，P = 0.028）、中位数（t = Z = 3.701，P = 0.000）、25% 分位数（t = -3.033，P = 0.003）和 75% 分位数（t = -3.441，P = 0.001）均高于低级别胶质瘤，FA 值（Z = 4.489，P = 0.000）、MK 值（Z = 4.528，P = 0.000）和MKT 值（Z = 4.528，P = 0.000）偏度均低于低级别胶质瘤。Logistic 回归分析显示，FA 值偏度（OR = 0.484，95%CI：0.278 ~ 0.842；P = 0.010）、MD 值方差（OR = 2.821，95%CI：1.231 ~ 6.466；P = 0.014）和 MKT 值 75% 分位数（OR = 2.581，95%CI：1.148 ~ 5.806；P = 0.022）是胶质瘤分级的预测因素。ROC 曲线显示，DKI 直方图联合 EphA2 评分的曲线下面积为0.90 ± 0.02（95%CI：0.676 ~ 0.922，P = 0.000），其诊断效能优于DKI 直方图的0.86 ± 0.02 （95%CI：0.809 ~ 0.916，P = 0.000；Z = 1.114，P = 0.041）。Spearman 秩相关分析显示，仅 MD 值峰度与 EphA2 评分呈负相关关系（r_s = -0.267，P = 0.002）。结论 DKI 直方图联合 EphA2 评分的胶质瘤分级预测模型可以有效提高胶质瘤分级诊断效能。

关键词: 神经胶质瘤, 膜蛋白质类, 弥散磁共振成像, 预测, Logistic 模型

Abstract: Objective To investigate the value of diffusion kurtosis imaging (DKI) histogram combined with Ephrin type-A receptor 2 (EphA2) in the evaluation of glioma grading. Methods A total of 183 patients with diffuse glioma [including 63 cases of low -grade glioma (LGG) and 120 cases of high-grade glioma (HGG)] who underwent neurosurgical resection and were confirmed by pathology at General Hospital of Eastern Theater Command from December 2015 to December 2019 were enrolled. All patients underwent conventional MRI and DKI examination [including fractional anisotropy (FA), mean diffusivity (MD), kurtosis fractional anisotropy (KFA), mean kurtosis (MK), mean kurtosis tensor (MKT)], and DKI histogram parameters (including mean, variance, median, 25% quantile, 75% quantile, skewness, kurtosis) were obtained. Immunohistochemical staining of EphA2 was performed. Univariate and multivariate Logistic regression analysis were used to screen the predictive factors of glioma grading, and based on these factors, the DKI histogram and the DKI histogram combined with EphA2 grading diagnostic prediction model were constructed, and the receiver operating characteristic curve (ROC) was drawn to evaluate its diagnostic efficiency. Spearman rank correlation analysis was used to explore the correlation between the DKI histogram parameters and the EphA2 grading. Results For HGG, the variance (t =-2.050, P = 0.042) and 75% quantile (t =-2.130, P = 0.035) of FA value, the variance (t =-6.052, P = 0.000) and skewness (Z =-3.326, P = 0.001) of MD value, the mean (t =-3.094, P = 0.002), variance (t =-2.228, P = 0.027), median (Z =-3.444, P = 0.001), 25% quantile (t =-3.022, P = 0.003) and 75% quantile (t =-3.438, P = 0.001) of MK value, the mean (t =-3.096, P = 0.002), variance (t =-2.140, P = 0.028), median (t =-3.701, P = 0.000), 25% quantile (t =-3.033, P = 0.003) and 75% quantile (t =-3.441, P = 0.000) of MKT value were higher than those of LGG. The FA value (Z = 4.489, P = 0.000), MK value (Z = 4.528, P = 0.000) and MKT value (Z = 4.528, P = 0.000) were significantly lower than those of LGG. Logistic regression analysis showed the skewness of FA value (OR = 0.484, 95%CI: 0.278-0.842; P = 0.010), variance of MD value (OR = 2.821, 95%CI: 1.231-6.466; P = 0.014) and 75% quantile of MKT value (OR = 2.581, 95%CI: 1.148-5.806; P = 0.022) were the predictive factors for glioma grading. The ROC curve showed the area under the curve (AUC) of DKI histogram parameters combined with EphA2 grading was 0.90 ±0.02 (95%CI: 0.676-0.922, P = 0.000), which was better than DKI histogram (0.86 ±0.02; 95%CI: 0.809-0.916, P = 0.000; Z = 1.114, P = 0.041). Spearman rank correlation analysis showed only MD kurtosis was negatively correlated with EphA2 grading (r_s =-0.267, P = 0.002). Conclusions The prediction model of DKI histogram combined with EphA2 grading can effectively improve the efficiency of grading diagnosis of glioma.

Key words: Glioma, Membrane proteins, Diffusion magnetic resonance imaging, Forecasting, Logistic models