CT平扫图像三维纹理分析对甲状腺良恶性结节的鉴别诊断价值

目的:探讨CT平扫图像三维纹理特征对甲状腺良恶性结节的诊断价值。方法:回顾性收集53例甲状腺良恶性结节患者资料，依据病理分为两组，其中良性结节组33例，恶性结节组20例。采用纹理分析专用软件对CT平扫图像进行三维纹理分析，提取整个病变的纹理特征， 通过软件自带的三种纹理特征提取方法(交互信息提取、Fisher系数提取和分类错误概率联合平均相关系数提取)各提取10个纹理特征，剔除重复特征。再通过共线性检测，剔除高共线性特征，共计8个纹理特征。根据Shapiro-Wilk检验，对正态分布的特征行两独立样本t检验，不符合正态分布的特征行Mann-Whitney U检验。对P＜0.05的特征绘制受试者工作特征曲线(ROC)并计算曲线下面积(AUC)，比较诊断效能。结果:提取的30个最具特征纹理特征中，Sigma和S(2,-2)AngScMom有统计学意义(P＜0.05),其余6个特征均不具有统计学意义(P均>0.05)。特征Sigma的AUC值为0.725，敏感度和特异度分别为60.00%和87.76%;特征S(2,-2)AngScMom的AUC 值为0.782，敏感度和特异度分别为80.00%、69.39%。Sigma和S(2,-2)AngScMom联合AUC 值为0.864，敏感度和特异度为66.67%、97.96%,诊断效能更优。结论:基于CT平扫的三维纹理分析技术对于甲状腺良恶性结节的鉴别诊断有一定的价值。

Three-dimensional texture analysis of unenhanced CT images in the differential diagnosis of benign and malignant thyroid nodules

Objective:To explore the diagnostic value of three-dimensional texture features on unenhanced CT images in differential diagnosis of benign and malignant thyroid nodules.Methods:The data of 53 patients with benign and malignant thyroid nodules were retrospectively analyzed,and the patients were divided into two groups based on pathological results,including 20 malignant nodules and 33 benign nodules.Special software for texture analysis was used to analyze the 3D texture on unenhanced CT to extract the texture features from the entire nodules.In addition,through the built-in software features extraction method,Fisher coefficient extraction,mutual information (MI) and classification error probability combined average correlation coefficient extraction (POE+ACC) extracted 10 texture features respectively,and excluding duplicate features.After collinearity detection,high collinearity features were excluded,resulting in eight texture features.Testing for group differences was performed by using Mann-Whitney U test or Student's t test after assessing the normality of data by using the Shapiro-Wilk test.The receiver operating characteristic curve (ROC) was plotted for the parameters with statistical differences,and the area under the curve (AUC) was calculated for diagnostic performance analysis.Results:From the 30 texture features extracted,Sigma and S (2,-2) AngScMom were statistically significant(P＜0.05).And other six features were not statistically significant (P＞0.05).The AUC of Sigma was 0.725,and the sensitivity and specificity were 60.00% and 87.76%,respectively.The AUC of the S (2,-2) AngScMom was 0.782,and the sensitivity and specificity were 80.00% and 69.39%,respectively.The combined AUC value of the two features was 0.864,and the sensitivity and specificity were 66.67% and 97.96%,respectively,and the diagnostic performance was better than the single feature.Conclusion:The 3D texture analysis technology based on unenhanced CT obtain satisfactory diagnosis performance of identifying benign and malignant thyroid nodules.