基于纹理特征的神经网络分类器用于肝硬化核磁共振图像分类识别

Use of texture-based neural network classifier in classified recognition of liver cirrhosis MRI

目的 研究基于纹理特征的神经网络分类器用于肝硬化核磁共振图像（MRI）分类诊断方法.方法 选取经大连医科大学附属第二医院临床和实验室检查确诊的18例患者的肝脏MR图像,其中肝硬化10例,正常肝脏8例,通过手工分割共获取MR图像感兴趣区（ROI）170个（肝硬化组88个,正常肝脏组82个）.通过灰度共生矩阵提取了2组170个ROI0°、45°、90°、135°4个方向的纹理特征参数（共计56个）,采用盒状图评估56个纹理特征参数区分肝硬化和正常肝脏的性能,获得2组间可分性好的纹理特征参数24个.分别采用全部的56个纹理特征参数（特征组A）、完全随机选择24个纹理特征参数（特征组B）及两组间可分性好的24个纹理特征参数（特征组C）训练反向传播（BP）神经网络,其中用于网络训练的ROI为110个,而测试BP神经网络的ROI为60个.结果 盒状图评价显示0°,45°,90°,135°4个方向上的能量、对比度、相关性、逆差矩、和方差以及差平均共计24个特征参数在肝硬化组和正常肝脏组间可分性较好.特征组C的正确识别率最高（95.00%,57/60）,高于特征组A和特征组B（78.33%,47/60;88.33%,53/60;P＜0.05）.结论 基于纹理特征的BP神经网络分类器适于肝硬化和正常肝脏MR图像的分类识别.

Objective To explore the use of neural network classifier as a method for classified MRI diagnosis of liver cirrhosis. Methods MR images of 18 patients with confirmed diagnosis based on clinical and laboratory investigations in the Second Affiliated Hospital of Dalian Medical University were included in the study. These patients comprised 10 with cirrhosis and 8 with normal liver. Manual segmentation of these MRIs yielded 170 regions of interest （ROIs） which included 88 of liver cirrhosis and 82 of normal liver. Each of 14 texture eigenvalues （56 in total） in four directions （0°, 45°, 90° and 135°） of these two groups of ROIs were extracted with grayscale co-occurrence matrices. Performance of the 56 texture eigenvalues in discrimination of cirrhotic from normal liver tissues was evaluated by box plot, thereby 24 of 56 texture eigenvalues with good discrimination performance were obtained. A neural network BP classifier was trained with the whole set of 56 eigenvalues （Eigen set A）, or a 24-value random subset （Eigen set B）,or a subset containing the 24 well-discriminating eigenvalues （Eigen set C）. Training set included 110 ROIs and the testing set involved 60 ROIs. Results The box plot evaluation revealed 24 texture eigenvalues in directions 0°, 45°, 90° and 135° （energy, resolution, relevance, sum-of-square difference, mean deviation,etc） that best discriminated MRI of cirrhotic liver from normal. Eigen set C yielded the highest rate of correct recognition （95.00%,57/60） as compared with eigen sets A （78.33%,47/60） and B （88.33%, 53/60, P〈0.05）.Conclusion The texture-based neural netword classifier seems ideal for classifying MR imaging of liver cirrhosis from normal.