论文部分内容阅读
目的:探讨MRI影像学特征联合纹理分析技术在术前预测肾透明细胞癌（ccRCC）WHO/国际泌尿病理学会（ISUP）核分级的价值。方法:回顾性分析2016年7月至2020年7月福建医科大学附属第一医院78例经手术病理确诊为ccRCC患者的术前肾脏MRI图像，根据WHO/ISUP分级系统，分为低级别组（49例，Ⅰ级2例，Ⅱ级47例）和高级别组（29例，Ⅲ级25例，Ⅳ级4例），并以随机数表法按照7∶3的比例将患者分配到训练集（n n=63）与验证集（n n=15）。评价MRI影像学特征并提取图像纹理特征。选取横断面图像病灶的最大层面，分别在Tn 2WI及皮髓质期（CMP）图像上勾画ROI。使用MaZda软件提取纹理特征，包括灰度直方图、灰度共生矩阵、游程矩阵、梯度、自回归模型及小波变换6种类型。联合Fisher法、分类错误概率结合平均相关系数、交互信息3种方法对提取的纹理特征进行初步筛选，并对筛选出的纹理参数或影像学特征进行两组间独立样本n t检验、Mann-Whitney U检验或χ2检验，差异有统计学意义的参数构建多因素二元logistic回归模型，建立ROC曲线，分析其术前预测ccRCC核分级的效能。n 结果:训练集中，低、高级别ccRCC组间肿瘤长径、形状与边界、CMP强化程度、静脉瘤栓和47个纹理特征差异有统计学意义。在训练集中，构建7个多因素二元logistic回归模型，包括影像学特征模型（M1）、Tn 2WI纹理特征模型（M2）、CMP图像纹理特征模型（M3）、影像学特征联合Tn 2WI纹理特征模型（M4）、影像学特征联合CMP图像纹理特征模型（M5）、Tn 2WI纹理特征联合CMP图像特征模型（M6）以及联合所有特征模型（M7）。其中M7预测ccRCC核分级的ROC曲线下面积最大，在训练集和验证集中，曲线下面积分别为0.901（95%CI 0.828~0.974）、0.820（95%CI 0.564~0.974）。n 结论:基于MRI纹理分析联合影像学特征有望成为术前无创预测ccRCC WHO/ISUP核分级的有效方法。“,”Objective:To explore the application value of MRI texture analysis in combination with imaging features to predict the WHO/International Society of Urological Pathology (ISUP) nuclear grading in pre-operative patients with clear cell renal carcinoma (ccRCC).Methods:MRI images of 78 patients diagnosed as ccRCC by surgical pathology from July 2016 to July 2020 in First Affiliated Hospital of Fujian Medical University were retrospectively analyzed. According to the WHO/ISUP grading system, the patients were divided into low grade group (49 cases, grade Ⅰ in 2 cases and grade Ⅱ in 47 cases) and high grade group (29 cases, grade Ⅲ in 25 cases and grade Ⅳ in 4 cases), and then were assigned to training set (n n= 63) and validation set (n n=15) in a ratio of 7∶3 using random indicator method. MRI radiological features were evaluated and MRI imaging texture features were extracted. The largest-diameter slice of lesion on cross-sectional images was selected and ROIs were drawn on Tn 2WI and corticomedullary phase (CMP) images, respectively. Quantitative texture analysis software MaZda was used to extract texture features, including gray-scale histogram, co-occurrence matrix, run-length matrix, gradient, autoregressive model and wavelet transform. The extracted texture features were preliminarily selected by the combination of Fisher, probability of classification errorand average correlation coefficient, and interaction information, and then the reduced texture parameters or imaging features were tested by the independent sample n t test, Mann-Whitney U test or χn 2 test. Parameters with statistically significant differences were used to construct a multi-factors binary logistic regression model and the ROC curve was used to analyze its effectiveness in predicting high grade ccRCC.n Results:In training set, there were significant differences intumor length, shape and margin, enhancement degree of CMP, vein thrombosis and 47 texture features between the low and high grade ccRCC groups. In the training set, 7 multi-factors binary logistic regression model were constructed, including radiological features model (M1), Tn 2WI texture features model (M2), CMP image texture features model (M3) and combination radiological features of Tn 2WI texture features model (M4), combination radiological features of CMP images texture features model (M5), combination Tn 2WI texture features of CMP images texture features model (M6) and combination of all features model (M7). The area under ROC curve of M7 in predicting nuclear grading of ccRCC was the largest, which were 0.901 (95% CI 0.828-0.974) and 0.820 (95% CI 0.564-0.974) in the training set and validation set, respectively.n Conclusion:MRI texture analysis combined with imaging features is hopeful to be an effective preoperative noninvasive method in predicting WHO/ISUP grading of ccRCC.