摘要
目的探讨基于增强CT的MaZda纹理分析技术预测肝细胞癌分化程度的价值及最佳效能分析。方法回顾性分析经手术病理证实的128例单发肝细胞癌患者的增强CT图像,其中高级别组60例,低级别组68例。使用MaZda软件的6种纹理提取方法(直方图、灰度共生矩阵、游程矩阵、梯度模型、自回归模型和小波转换)及4种筛选方法(Fisher、POE联合ACC、MI和FPM),提取并筛选各期肿瘤纹理特征。使用B11软件的4种方法(RDA、PCA、LDA和NDA)对病灶进行分类,比较增强各期相、不同统计方法组合下纹理特征集预测肝细胞癌病理分级的最小误判率。使用二元Logistic回归建立模型,采用ROC曲线评估纹理特征及模型的预测价值。结果在众多联合方法中,增强各期FPM联合NDA产生的误判率均低于其他方法,其中门脉期误判率最低3.13%,动脉期10.94%,延迟期7.81%,均在良好范围内。门脉期FPM筛选的30个纹理特征中23个有统计学意义(P<0.05),纳入二元Logistic回归模型筛选出3个纹理特征:Perc.99、S(3,-3)Contrast及WavEnLL_s-3。回归模型的曲线下面积为0.915,敏感度100%,特异度70%。结论基于常规增强CT的MaZda纹理分析技术能够预测肝细胞癌的分化程度;不同增强时期及不同筛选分类方法相组合会产生不同的预测效能,其中以门脉期联合FPM及NDA为佳。门脉期联合FPM获得的纹理特征经过Logistic回归建模,模型诊断效能较高。
Objective To investigate the efficacy of MaZda texture analysis on enhanced CT in predicting the pathological grade of hepatocellular carcinoma(HCC).Methods 128 patients with pathologically confirmed solitary high-grade(60)or low-grade(68)HCC underwent conventional contrast-enhanced CT.MaZda software was used to compute the texture features of the HCC including histogram-,co-occurrence matrix-,run-length matrix-,gradient-map-,autoregressive model-and Haar wavelet transform-based features.Then 4 statistical methods including Fisher's coefficient(F),probability of classification error and average correction coefficient(POE+ACC),mutual information(MI),and a combination of the 3 methods(FPM)in the MaZda software were used to select the optimal texture features for differentiating high-grade and low-grade HCC.Principal component analysis(PCA),linear discriminant analysis(LDA),nonlinear discriminant analysis(NDA),and raw data analysis(RDA)in the B11 module of MaZda were used to reduce dimensionality and classify these texture parameters.The result was expressed as misclassification rate in predicting pathological grading of HCC and was compared among the different phases of contrast enhancement and various statistical methods.The texture features with the minimum misclassification rate were compared between the high-grade and low-grade HCC.Binary logistic regression was used to establish the model.The predictive value of the texture features and model was assessed using the receiver operating characteristic curve.Results The minimum misdiagnosis rates of FPM+NDA in the arterial(10.94%),venous(3.13%),and delayed(7.81%)phases were lower than that of other methods.Among the 30 textural features screened by FPM in the venous phase,23 were statistically significant(P<0.05).Three texture features,perc.99,s(3,-3)contrast and WavEnLL_s-3 were selected by the binary logistic regression model with the area under the ROC curve of 0.915,sensitivity of 100%,and specificity of 70%.Conclusion Texture analysis on enhanced CT can help predicting the pathological grading of HCC.The combination of FPM texture feature selection method and NDA dimensional reduction method in the venous phase has highest diagnostic efficiency.
作者
张月
黄顺根
蒋震
盛茂
郭万亮
ZHANG Yue;HUANG Shun-gen;JIANG Zhen;SHENG Mao;GUO Wan-liang(Department of Radiology,Children’s Hospital of Soochow University,Jiangsu 215006,China)
出处
《影像诊断与介入放射学》
2022年第3期163-168,共6页
Diagnostic Imaging & Interventional Radiology
基金
江苏省卫生健康委科研项目(M2020068)。
关键词
肝细胞癌
病理分化程度
体层成像术
X线计算机
纹理分析
MaZda软件
Hepatocellular carcinoma
Pathological differentiation degrees
Tomography,X-ray computed
Texture analysis
MaZda software
