摘要目的 探讨增强MRI的纹理分析技术术前预测肝细胞癌(HCC)微血管侵犯(MVI)的价值.方法 回顾性分析中国医学科学院肿瘤医院2014年1月至2016年12月经手术病理证实的HCC患者60例,据手术病理结果分为MVI组30例,无MVI组30例.患者术前均行肝脏常规MRI平扫及动态对比增强MRI检查.采用美国GE Omni-Kinetics软件对DCE-MRI原始数据分别在动脉期和门静脉期进行纹理特征提取67个纹理特征,采用独立样本t检验筛选MVI组和无MVI组间差异有统计学意义的特征.增强MRI图像纹理特征采用单纯主成分分析(PCA)和建模(降维、建模、预测与验证)两种方法.采用logistic回归方法建立模型,以组织病理诊断为金标准,把动脉期和门静脉期数据的80%作为训练组(48例,MVI和无MVI组各24例),20%作为验证组(12例,MVI和无MVI组各6例),分别对动脉期和门静脉期增强图像进行建模和交叉验证,并采用ROC评价模型的诊断效能.结果 MVI组和无MVI组间差异有统计学意义的动脉期纹理有15个,门静脉期纹理3个.PCA法提取重要的MRI纹理特征,对动脉期15个特征进行相关分析,发现归一化正像素能量值与能量之间具有很好的相关(r>0.90),故删除归一化正像素能量值.然后进行14个特征的PCA统计分析,发现重要的纹理特征参数有4个,分别为灰度共生矩阵相关性、Hara方差、灰度共生矩阵方差和及灰度共生矩阵熵和.门静脉期重要的纹理特征参数有3个,分别是病灶的灰度共生矩阵差分熵,其次是长游程低灰阶显著性和灰度共生矩阵差分方差.建立模型、预测交叉验证,动脉期提取的3个特征分别为灰度共生矩阵相关性、灰度共生矩阵对比度及灰度共生矩阵熵和,门静脉期提取2个特征分别为灰度共生矩阵差分方差和长游程低灰阶显著性.训练组动脉期、门静脉期模型,验证组动脉期、门静脉期模型诊断MVI的ROC下面积分别为0.774、0.681、0.889、0.611,动脉期诊断准确率(83.30%,10/12)高于静脉期(42.00%,5/12).结论 采用DCE-MRI图像纹理分析技术能够在HCC术前预测MVI,动脉期纹理特征的预测准确度更高.

abstractsObjective To investigate the prognostic value of the texture analysis contrast-enhanced MR imaging (DCE-MRI) in predicting microvascular invasion in hepatocellular carcinoma (HCC) before operation. Methods Sixty patients with HCC confirmed by pathology in the Chinese Academy Medical Sciences from January 2014 to December 2016,were enrolled in our study retrospectively.According to the post-operative pathology, the patients were divided into positive microvascular invation[MVI(+)]group including 30 patients, and negative MVI[MVI(-)] group including 30 patients. All patients underwent normal MR and DCE-MRI before surgery.Sixty seven texture features were extracted from the original data of arterial phase (AP) and portal venous phase (PVP) of DCE-MRI. All data were calculated by using Omni-Kinetics(OK)software of the United States.The difference between MVI(+)group and MVI(-)group was statistically significant using the independent sample t test. The identified methods of the DCE-MR texture features in predicting MVI adopted the principal component analysis (PCA) and the establishing prediction model including dimensionality reduction, modeling, prediction and verification. The model was established by logistic regression method. According to the histopathology, 80% data of AP and PVP were used as training group[48 cases,MVI(+)and MVI(-)group 24 cases respectively],20% as validation group [12 cases, MVI(+) and MVI(-) group 6 cases respectively]. The DCE-MRI images of AP and PVP were modeled and cross-referenced respectively, and the diagnostic efficiency of ROC evaluation model was adopted. Results There were 15 significant different texture features of the AP and three significant different texture features of the PVP between MVI(+) group and MVI(-) group respectively. The PCA method extracted the important DCE-MRI texture features and analyzed the 15 features of AP.The UPP and energy showed a good correlation(r>0.90),therefore the UPP were removed.Fourteen texture features were analyzed using the PCA method.There were four important texture features including the GLCM Correlation, Hara Variance, GLCM sum Variance and GLCM sum Entropy in the AP. Moreover, there were three important texture features including GLCM difference Entropy, Long Run Low Grey Level Emphasis and GLCM difference Variance in the PVP.Through the prediction model was established and crossly validated. There were three significant different texture features in the AP of DCE-MRI,including GLCM Correlation, GLCM Contrast and GLCM sum Entropy.And there were two significant different texture features in the PVP of DCE-MRI,including GLCM difference Variance and Long Run Low Grey Level Emphasis.In the training and validation group,the areas under the ROC of the AP model and PVP model were 0.774,0.681,0.889 and 0.611 respectively.The diagnosed accuracy rate of the AP model(83.30%,10/12)was higher than that of the PVP model(42.00%,5/12).Conclusion The DCE-MRI texture analysis technique could predict the MVI of HCC before operation,and the predictive accuracy of the AP texture feature was higher.