MRI纹理分析鉴别诊断肝硬化背景下小肝癌与增生结节

MRI texture analysis in differential diagnosis of small hepatocellular carcinoma and dysplastic nodules in cirrhosis liver

目的探讨常规MRI纹理分析鉴别诊断肝硬化背景下小肝癌与增生结节的价值。方法回顾分析经病理证实的33例小肝癌和19例肝增生结节患者的MRI资料。采用MaZda软件手工勾画ROI,提取T1WI、T2WI、频率选择性预脉冲脂肪抑制T2WI及T1WI增强扫描图像中病变的纹理特征。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)及三者联合(FPM)的方法选择最佳纹理参数集合。使用原始数据分析(RDA)、主要成分分析(PCA)、线性判别分析(LDA)和非线性判别分析(NDA)进行纹理分类。同时由2名MRI诊断医师共同评估所有影像学资料。比较纹理分析与医师鉴别诊断两种病变结果的差异。结果 52例中,共60个病灶。鉴别小肝癌与增生结节的纹理特征主要来自T2WI,误判率最小为8.33%(5/60)。纹理特征选择方法中,FPM的误判率(8.33%～26.67%)均低于MI(20.00%～38.33%)、Fisher(18.33%～41.67%)和POE+ACC(8.33%～40.00%)。纹理特征分类方法中,NDA判别两种病变的误判率(8.33%～20.00%)均低于RDA(26.67%～41.67%)、PCA(28.33%～43.33%)和LDA(21.67%～45.00%)。影像医师的误判率为23.33%(14/60),高于采用纹理分析鉴别两种病变的误判率(5/60,8.33%;χ~2=58.73,P=0.002)。结论常规MRI纹理分析可用于鉴别肝硬化背景下小肝癌与增生结节。

Objective To investigate the value of texture analysis of conventional MRI in differential diagnosis of small hepatocellular carcinoma（sHCC）and dysplastic nodules（DN）in cirrhotic liver.Methods MRI data of 33 patients with sHCC and 19 patients with DN proven pathologically were retrospectively analyzed.The texture features of lesions based on axial T1WI,T2WI,spectral attenuated inversion-recovery T2WI and contrast-enhanced T1WI were extracted by using manually drawing ROIs with software MaZda.The subsets of optimized texture parameters were chosen with four different methods,i.e.Fisher coefficient,the probability of classification error and average correlation（POE＋ ACC）,mutual information measure（MI）and combination of the above three methods（FPM）,respectively.Raw data analysis（RDA）,principal component analysis（PCA）,linear discriminant analysis（LDA）and nonlinear discriminant analysis（NDA）were performed for texture classification.Two radiologists with MRI diagnostic experience were requested to evaluate all the images.The differences between texture analysis results and physician＇s diagnostic results were compared.Results Among 52 patients,60 lesions were found.The texture features in differential diagnosis of sHCC and DN in cirrhotic liver were mainly from T2WI,which had the lowest misclassification rate（8.33% [5/60]）.Among the texture feature selection methods,the misclassification rate of FPM（8.33%—26.67%）was lower than that of MI（20.00%—38.33%）,Fisher（18.33%—41.67%）and POE＋ACC（8.33%—40.00%）.The misclassification rate of NDA（8.33%—20.00%）was lower than that of RDA（26.67%—41.67%）,PCA（28.33%—43.33%）and LDA（21.67%—45.00%）. The misclassification rate of radiologists diagnosis was 23.33%（14/60）,higher than that of texture analysis（5/60,8.33%;χ^2=58.73,P=0.002）.Conclusion It is feasible to use texture analysis on conventional MRI for differentiation of sHCC and DN in cirrhotic liver.