CT平扫图像纹理分析技术在肺部包块诊断中的应用  

Application of texture analysis based on plain CT image in the diagnosis of pulmonary mass

作　　者：张志明 王德银 ZHANG Zhiming;WANG Deyin(Department of Radiology,Pinghu First People’s Hospital,Pinghu,Zhejiang Province 314200,China)

机构地区：[1]平湖市第一人民医院放射科,浙江平湖314200

出　　处：《实用放射学杂志》2022年第12期1951-1953,1981,共4页Journal of Practical Radiology

摘　　要：目的探讨CT平扫图像纹理分析技术在肺部包块(非钙化肺结核球和浸润性肺腺癌)鉴别诊断中的应用价值。方法回顾性分析肺内包块患者60例的CT图像资料。使用MaZda纹理提取分析软件从各自组中获得300个纹理特征参数,采用Fisher参数法、最小平均相关系数(ACC)与最小分类误差法(POE)以及相关信息测度法(MI)3种方法进行最佳纹理特征的提取,获得更加具有意义的联合特征(MPF)。再通过MaZda软件对获得的4组最佳纹理特征进行线性判别分析法(LDA)和非线性判别分析法(NDA)分类,计算各组最佳纹理特征鉴别2种疾病时出现的最小错误率,通过受试者工作特征(ROC)曲线比较其诊断效能。结果单种方法中NDA/人工神经网络(ANN)-Fisher错误率最低,仅为6.67%,MPF中NDA/ANN-MPF错误率最低,仅为5.00%,两者比较差异较小。2种疾病之间最佳纹理特征间存在10个差异有统计学意义特征,通过ROC曲线进行诊断效能比较显示差方差S(1,1)、差异熵S(1,1)以及梯度方差诊断效能较高,且三者间曲线下面积(AUC)比较差异无统计学意义(P>0.05)。结论CT平扫图像纹理特征分析可较好地区分浸润性肺腺癌和非钙化肺结核球,为鉴别诊断提供可靠的客观依据。Objective To explore the application value of texture analysis based on plain CT image in the differential diagnosis of pulmonary mass(non-calcified pulmonary tuberculosis and invasive lung adenocarcinoma).Methods The CT image data of 60 patients with pulmonary mass were analyzed retrospectively.A MaZda texture extraction and analysis software was used to obtain 300 texture feature parameters from the CT image of patients.Fisher parameter method,minimum average correlation coefficient(ACC),minimization of probability of classification error(POE)and mutual information coefficients(MI)were used to extract the best texture features,and thereafter 3 sets of texture feature parameters were obtained.Finally,3 sets of parameters were combined to obtain the 4 set of parameters,named get more meaningful joint characteristics(MPF).MaZda software was uesd to classify the 4 sets of best texture features by linear discriminant analysis(LDA)and nonlinear discriminant analysis(NDA),and to calculate the minimum error rate of each set best text features in identifying two kinds of diseases,and then the diagnostic efficacy was compared through the receiver operating characteristic(ROC)curve.Results NDA/artificial neural network(ANN)-Fisher got the lowest error rate of 6.67%,and the NDA/ANN-MPF had the lowest error rate in MPF of 5.00%.There were 10 best texture features with statistically significant differences in the two kinds of diseases.ROC curve analysis showed high diagnostic efficacy for the difference variance S(1,1),the difference entropy S(1,1)and the gradient variance diagnostic power,and had no statistically significant difference in area under the curve(AUC)between these parameters(P>0.05).Conclusion The texture feature analysis based on plain CT image can better distinguish invasive lung adenocarcinoma from non-calcified pulmonary tuberculosis,and provides a reliable objective basis for differential diagnosis.

关 键 词：浸润性肺腺癌 非钙化肺结核球 纹理分析 计算机体层成像 

分 类 号：R734.2[医药卫生—肿瘤] R814.42[医药卫生—临床医学]