基于纹理特征的多序列MRI的肝硬化识别研究

目的 采用基于纹理特征的十倍交叉验证法的神经网络分类器，探讨多序列MRI在肝硬化诊断识别中的价值。方法 将5个序列MR图像（T1WI、T2WI、增强动脉期、门静脉期和平衡期）分成正常肝脏组、较早期肝硬化组及中晚期肝硬化组，手动获取ROI；采用灰度共生矩阵提取ROI的56个纹理特征参数；采用十倍交叉验证法的BP网络分类器分别分类识别3组肝脏组织。结果 门静脉期图像对正常肝脏、较早期肝硬化及中晚期肝硬化的分类效果最好，正确率为87.62%（92/105），T2WI分类效果最差，正确率为78.33%（47/60），T1WI、动脉期和平衡期图像居两者之间。结论 采用基于纹理特征的十倍交叉验证法的神经网络分类器可以区分正常肝脏、较早期和中晚期肝硬化MRI；在基于多序列MRI的肝硬化识别研究中，门静脉期图像有可能成为首选。

Multi-sequence MRI classification of hepatic cirrhosis based on texture feature

Objective To investigate the diagnostic value of multi-sequence dynamic MRI for hepatic cirrhosis using tenfold cross-validation method neural network classifier based on texture feature. Methods T1WI, T2WI, arterial phase, portal venous phase and equilibrium phase imaging were divided into normal, early and advanced stage hepatic cirrhosis groups. ROI of these images were cut manually. Fifty-six texture features were extracted by grey level co-occurrence matrices. Hepatic tissues were classified by a BP classifier based on tenfold cross-validation method. Results For classification of hepatic tissue in all 3 groups, imaging of portal venous phase were the best, and the total accuracy was 87.62% (92/105), T2WI were the worst, with the total accuracy of 78.33% (47/60). T1WI, imaging of equilibrium phase and arterial phase were all better than T2WI. Conclusion Tenfold cross-validation method neural network classifier can classify normal, early and advanced stage hepatic cirrhosis on MRI based on texture feature. Portal venous phase imaging may be the first choice for classification of hepatic cirrhosis based on multi-sequence MRI.