摘要
研究描述超声肝图像纹理特征的分维方法。用14幅正常肝样本图像和14幅原发性肝癌样本图像检验并比较评估4种分维方法。用布朗运动方法、毯子法、傅里叶功率谱法和差分盒计数法4种方法得到的分维作为特征进行ROC(Receiver operating characteristic)分析,以SVM作为模式分类方法的分类正确率进行分析。研究结果表明:除了分数布朗运动方法外,由毯子法、傅里叶功率谱法和差分盒计数法获得的描述正常肝图像感兴趣区域的分维值明显小于描述原发性癌图像感兴趣区域的分维值;采用傅里叶功率谱方法得到最大的ROC曲线下的面积;用SVM(Support vector machine)方法进行分类也取得了与ROC分析类似的结果,即用傅里叶功率谱方法进行分类准确度最高,分数布朗运动和差分盒计数方法效果较差,毯子方法效果居中;傅里叶功率谱方法是描述超声肝图像纹理特征最适合的方法。
A detailed study of fractal-based methods for texture characterization of normal ultrasonic liver parenchyma and primary liver cancer image was made.The 4 types of methods of fractal dimension estimation for the texture feature characterization of normal liver and primary liver cancer image were tested and compared and evaluated based on 14 normal liver and 14 primary liver cancer sample images.ROC was used as an analysis method and SVM was used as a pattern classification method for the performance evaluation of 4 types of methods using the fractal dimensions derived from 4 methods as the features.The results show that,except the fractional Brownian motion method,the fractal dimension of region of interest depicting normal liver parenchyma is statistically significantly lower than that of primary liver cancer with all other methods.The highest area below ROC curve was obtained using Fourier power spectrum.The accuracy results obtained by SVM analysis are similar to those obtained by ROC.Namely,the accuracy with Fourier power spectrum method is the highest,yet the accuracy for fractional Brownian motion and differential box counting method was lower,and the accuracy for Blanket method is intermediate.Fourier power spectrum method is the most appropriate method of characterizing the texture feature of ultrasonic liver image.
出处
《中南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2011年第9期2746-2753,共8页
Journal of Central South University:Science and Technology
关键词
超声肝图像
分维方法
傅里叶功率谱
纹理特征
ROC分析
SVM
ultrasonic liver image
fractal dimension method
Fourier power spectrum
texture feature
ROC analysis
SVM
