基于CT增强影像组学模型预测甲状腺乳头状癌颈部淋巴结转移

Preliminary study on radiomics based on CT contrast-enhanced images in predicting cervical lymph nodemetastasis in patients with papillary thyroid carcinoma

目的:探讨基于CT增强图像的组学特征预测甲状腺乳头状癌(PTC)颈部淋巴结转移的可行性。方法:回顾性分析130个经病理证实的PTC结节(伴颈部淋巴结转移65个,无淋巴结转移65个)的CT增强扫描图像,应用Mazda纹理分析软件对CT图像进行预处理,并提取PTC结节的一阶和高阶纹理特征。应用Fisher相关系数、分类误差概率与平均相关系数(POE+ACC)和交互信息(MI)三种降维方法分别对纹理特征进行筛选,每种方法筛选出10个最佳纹理特征。应用Mazda软件的B11模块中的原始数据分析(RDA)、主成分分析(PCA)、线性判别分析(LDA)和非线性判别分析(NDA)四种统计学方法分别进行数据的进一步降维,计算不同统计学方法、不同降维方法组合下建立的影像组学模型的错判率、敏感度和特异度。结果:最佳纹理参数主要来源于二阶及高阶纹理特征(灰度游程矩阵、灰度共生矩阵和小波转换)。不同降维方法组合同一统计学方法时所获模型的错判率较为接近,范围为3.08%~16.92%。三种降维方法分别与非线性判别分析(NDA)组合时,所获模型的诊断结果一致,与另外3种统计学方法相比敏感度和特异度最高,分别为98.46%、95.38%,错判率最低(3.08%)且差异均有统计学意义(P<0.05)。结论:基于CT增强图像的PTC结节的二阶及高阶纹理特征对预测颈部淋巴结转移是可行的,不同的降维方法和统计学方法组合获得的组学模型诊断效能不同,推荐使用非线性判别分析与任一降维方法组合的组学模型。

Objective:To explore the feasibility of radiomics features based on contrast-enhanced CT images in predicting the cervical lymph node metastasis in patients with papillary thyroid carcinoma(PTC).Methods:The contrast-enhanced CT images of 130 patients with PTC(65 with CLNM and 65 without CLNM)confirmed by pathology were analyzed retrospectively.Texture analysis software(Mazda)was used to extract the texture features of the thyroid nodules in the two groups.Then the extracted texture parameters were dimensionally reduced,and the 10 optimal texture features were selected by Fisher coefficient,classification error probability with average correlation coefficients(POE+ACC)and mutual information coefficient(MI)in Mazda software.Then four statistic methods were used to further reduce the dimensionality respectively by using B11 software in Mazda software.The misdiagnosis rates of radiomics models from different combinations of statistic methods with dimensionality reduction methods were calculated.Results:The ten optimal texture features were all second-order or high-order texture features(from gray-level run matrix,gray-level co-occurrence matrix or wavelet transform).The misdiagnosis rates selected out by three dimensionality reduction methods with the same statistic method were closed to each other,ranged from 3.08%to 16.92%;NDA showed the lowest misdiagnosis rates(3.08%)and the highest sensitivity and specificity(98.46%and 95.38%,respectively)with statistically significant difference(P<0.05)when compared with the other three statistic methods.Conclusion:It is feasible for texture analysis in predicting the CLNM in PTC patients with second-order or high-order texture features.The different combinations of statistic methods and dimensionality reduction methods have different efficacy.NAD combined with either dimensionality reduction method is recommended,which has the lowest misdiagnosis rate and the hig-hest sensitivity and specificity.