基于CT增强扫描纹理分析术前预测胃腺癌病理分化程度的价值

Value of texture analysis based on enhanced CT in preoperative prediction of gastric adenocarcinoma pathological differentiation

目的:探讨CT增强扫描纹理分析术前预测胃腺癌病理分化程度的价值。方法:回顾性分析经手术病理证实的85例胃腺癌患者的临床、病理及CT资料。术后病理示低分化腺癌52例(低分化组),中高分化腺癌33例(中高分化组)。选取肿瘤最大层面的轴位CT门静脉期图像,采用MaZda软件沿肿瘤轮廓绘制ROI,提取胃癌纹理特征。纹理特征选择采用分类错误概率联合平均相关系数(POE+ACC)、交互信息(MI)及Fisher系数联合的方法,筛选30个与胃癌病理分化相关的纹理特征。采用4种分类方法评价纹理特征判别胃癌病理分化的性能,包括主要成分分析(PCA)、原始数据分析(RDA)、线性分类分析(LDA)及非线性分类分析(NDA),纹理分类结果以误判率表示,采用ROC曲线分析预测效能。结果:LDA、PCA、RDA及NDA分类方法判别胃癌病理分化程度的总体错判率分别为14.12%、31.76%、25.88%、9.41%,ROC曲线的AUC分别为0.851、0.668、0.733、0.909;其中NDA的预测效能最佳,预测胃癌分化程度的敏感度为91.91%、特异度为90.38%、准确率为90.59%。结论:CT纹理分析能为胃腺癌病理分化程度术前预测提供无创、可靠方法,其中NDA分类方法的预测效能最优。

Objective:To investigate the value of contrast-enhanced CT texture analysis in predicting the degree of pathological differentiation of gastric adenocarcinoma.Methods:The clinical,pathological and preoperative CT imaging data of 85 gastric adenocarcinoma patients confirmed by operation and pathology were retrospectively analyzed.According to the postoperative pathological results,52 patients had poorly differentiated gastric adenocarcinoma,and 33 patients had moderately/well differentiated gastric adenocarcinoma.The axial CT portal phase image of the largest slice of the tumor was selected,an ROI was drawn along the tumor contour by MaZda software,and the texture features of gastric cancer were extracted.The combination method of the probability of classification error combined with the average correlation coefficient(POE+ACC),interactive information(MI)and the Fisher coefficient was used to select texture features,and a total of 30 texture features closely related to the pathological differentiation of gastric cancer were screened.Four classification methods,including principal component analysis(PCA),raw data analysis(RDA),linear classification analysis(LDA)and nonlinear classification analysis(NDA),were used to evaluate the performance of texture features in discriminating the pathological differentiation of gastric cancer.The results of texture classification were expressed by the misjudgment rate,and the prediction efficiency was evaluated by ROC curve analysis.Results:The overall misjudgment rate of LDA,PCA,RDA and NDA classification method was 14.12%,31.76%,25.88%,9.41%,respectively,and the AUC of ROC curve was 0.851,0.668,0.733,0.909,respectively.Among them,NDA showed the best efficiency,with a sensitivity of 91.91%,specificity of 90.38%and accuracy of 90.59%in predicting the pathological differentiation degree.Conclusion:CT texture analysis can provide a noninvasive and reliable method for the preoperative prediction of the degree of pathological differentiation of gastric adenocarcinoma,among which the NDA method shows the best efficiency.