CT纹理分析鉴别结肠癌肠周脂肪癌性侵犯的价值  

The value of CT texture analysis in differentiating cancerous invasion of pericolic fat in colorectal cancer

作　　者：王晗 陈雨清 郭盛仁[1] 陈光强[1] WANG Han;CHEN Yuqing;GUO Shengren;CHEN Guangqiang(Department of Imaging Diagnosis,the Second Affiliated Hospital of Soochow University,Suzhou,Jiangsu Province 215008,China;Department of Ultrasound,the Affiliated Hospital of Xuzhou Medical University,Suzhou,Jiangsu Province 221000,China)

机构地区：[1]苏州大学附属第二医院影像诊断科,江苏苏州215008 [2]徐州医科大学附属医院超声科,江苏徐州221000

出　　处：《实用放射学杂志》2020年第10期1601-1604,1614,共5页Journal of Practical Radiology

摘　　要：目的探讨常规CT纹理分析鉴别结肠癌肠周脂肪癌性侵犯的应用价值。方法选取结肠癌患者87例,按病理结果分2组:肠周脂肪癌性侵犯42例,无癌性侵犯45例。MaZda软件提取肠周脂肪纹理特征,交互信息(MI)、Fisher系数、分类错误概率联合平均相关系数(POE+ACC)、上述方法联合(FPM)筛选纹理特征,使用原始数据分析(RDA)、主成分分析(PCA)、线性分类分析(LDA)、非线性分类分析(NDA)4种分类分析统计方法对2组病变分类,错判率(MCR)评估鉴别诊断效果。以获最小MCR的分类序列所筛纹理特征建立Logistic回归模型,降维得出最佳纹理特征。结果动脉期纹理特征最有意义;FPM所筛纹理特征经分类后所获最小MCR次数最多;NDA分类2种病变的MCR均最低。最小MCR为3.45%,由FPM选择出30个动脉期纹理特征后经NDA分类所得,将这30个纹理特征纳入Logistic回归模型,准确率为93.1%,敏感度为91.1%,特异度为95.2%,AUC为0.967,降维得到99%百分位数(Perc.99%)、熵(Entropy)及小波能量(WavEnHL_s-2)鉴别意义较大。结论常规CT纹理分析可用于鉴别结肠癌肠周脂肪有无癌性侵犯,提高分期准确率,为患者术前治疗方案的选择提供更有意义的参考价值。Objective To explore the application value of conventional CT texture analysis in differentiating cancerous invasion of pericolic fat in colorectal cancer.Methods 87 patients with colorectal cancer were analyzed,including 42 cases with cancerous invasion of pericolic fat and 45 without cancerous invasion.MaZda software was used to extract texture features of pericolic fat.Mutual information(MI),Fisher coefficient,classification error probability combined average correlation coefficients(POE+ACC)and the combination of the three methods(FPM)were used to screen out the texture features.The 2 groups of lesions were classified by four types of statistical analysis methods:raw data analysis(RDA),principal component analysis(PCA),linear discriminant analysis(LDA)and nonlinear discriminant analysis(NDA).The effect of differential diagnosis was rated by misclassification rate(MCR).A Logistic regression model was built by the texture features selected based on the classification sequence with the minimum MCR.Furthermore,the best texture features were obtained by reducing the dimensions of the Logistic regression model.Results Arterial phase texture features were the most significant.Texture features screened by FPM obtained the most frequently MCR.NDA had the minimum MCR in differentiating the two lesions.The minimum MCR was 3.45% obtained by NDA after 30 arterial phase texture features were selected by FPM.The 30 texture features were used to establish a Logistic regression model.The accuracy,sensitivity,specificity and AUC of the model were 93.1%,91.1%,95.2% and 0.967,respectively.Perc.99%,Entropy and WavEnHL_s-2 had great implications for differentiation by dimensionality reduction.Conclusion The quantitative information provided by conventional CT texture analysis can be used for differentiating cancerous invasion of pericolic fat in colorectal cancer,improving the accuracy of staging of colorectal cancer and providing more significant reference value for the selection of preoperative treatment scheme for patients.

关 键 词：结肠癌 计算机体层成像 T分期 纹理分析 

分 类 号：R735.3[医药卫生—肿瘤]