增强CT纹理分析对食管鳞癌转移性淋巴结的诊断价值

摘要

目的:探讨增强CT图像纹理分析对食管鳞癌转移性淋巴结的诊断价值.方法:回顾性分析48例行食管癌切除加淋巴结清扫患者的术前增强CT图像, 在CT增强图像上选取短径>5mm的食管癌区域淋巴结, 根据病理结果分为转移性淋巴结 (MLN) 和非转移性淋巴结 (NLN) 组.应用MaZda软件手动选出感兴趣区 (ROI) , 提取的纹理特征根据是否纳入几何参数分为两组, 分别通过Fisher系数、分类错误概率联合平均相关系数 (POE+ACC) 、交互信息 (MI) 及上述3种方法联合法 (FPM) 筛选出最具有鉴别MLN及NLN价值的纹理特征, 然后分别采用原始数据分析 (RDA) 、主要成分分析 (PCA) 、线性分类分析 (LDA) 和非线性分类分析 (NDA) 四种特征分类统计方法进行判断, 结果以错判率形式表示.结果:纳入几何参数时, 最低错判率为7.84％ (8/102) , 出现在特征选择方法采用POE+ACC、MI, 特征分类统计方法采用NDA时.不纳入几何参数时, 最低错判率为6.86％ (7/102) , 出现在特征选择方法采用FPM, 特征分类统计方法采用NDA时, 两者差异无统计学意义 (χ2=0.082, P=0.774).影像医师的错判率为14.71％ (15/102) , 较采用纹理分析鉴别两种病变的最低错判率高, 两者差异有统计学意义 (χ2=4.300, P=0.038).结论:增强CT纹理分析有助于鉴别食管鳞癌MLN与NLN, 为鉴别两者提供可靠的客观依据, 是否纳入几何参数对鉴别的结果无影响.%Purpose: To evaluate the diagnostic value of texture analysis on enhanced CT images in metastatic lymph nodes of esophageal squamous cell carcinoma. Methods: The preoperative CT images of 48 patients with esophageal carcinoma were retrospectively analyzed. According to the pathological results, the regional lymph nodes with a diameter of more than 5 mm were selected and divided into metastatic lymph nodes (MLN) and non-metastatic lymph nodes (NLN) groups. Firstly, ROI was manually delineated by MaZda software, and texture feature extraction was divided into two groups according to geometric parameters. The most discriminant character of MLN and NLN were selected by Fisher coefficient, POE + ACC, MI and FPM, respectively. The values of texture features were classified by four statistical methods, namely, original data analysis (RDA) , principal component analysis (PCA) , linear classification analysis (LDA) and nonlinear classification analysis (NDA). The results were expressed in the form of misjudgment rate. Results: The lowest misjudgment rate was 7.84％ (8/102) when geometric parameters were included. The feature selection method was POE+ACC, MI, and the feature classification and statistics method was NDA. When geometric parameters were not included, the lowest misjudgment rate was 6.86％ (7/102) , which appeared in the feature selection method using FPM, and feature classification statistical method using NDA. There was no significant difference between two methods (χ2= 0.082, P= 0.774). The misjudgment rate of radiologists was14.71％ (15/102) , which was higher than the lowest misjudgment rate of texture analysis and both was significant statistical difference (χ2= 4.300, P= 0.038). Conclusion: Texture analysis of enhanced CT is helpful to differentiate MLN from NLN in esophageal squamous cell carcinoma, and provides reliable objective basis for differentiating MLN from NLN.