【作者】 季桂树；

【导师】 何继善；

【作者基本信息】 中南大学 ， 生物医学工程， 2013， 博士

【摘要】 肝癌是危害人类健康甚至危及人类生命的主要病症,及早确诊这种病症是当务之急。随着计算机与图像处理技术的快速发展,超声等医学影像已成为对肝病进行诊断的主要信息源,同时也促进了模型方法在包括医学图像在内的图像特征提取方面的显著进步,这将更好地弥补用目视诊断方法远远不能满足对诊断结果要求的缺陷。因此成为了学界研究的热点。本文以正常肝和肝癌超声图像为数据样本,以肝脏超声图像纹理的分形特征为信息对肝脏超声图像计算机辅助诊断技术中的分类识别问题进行了深入的研究。所研究的内容包括：(1)分维和孔隙度的统计分析。对比分析了在其它领域对纹理图像进行分析均有上佳表现的4种分维方法,即毯子法(blanket)、傅立叶功率谱法(Fourier power spectrum, FPS)、分数布朗运动法(fractional Brownian motion, FBM)和差分盒计数法(differential box counting, DBC)。与此同时,又对分形几何学中的另一个重要参数——孔隙度的5种计算方法,即曼德尔布罗特法(Mandelbrot)、差分盒计数法(lacunarity of differential box counting, LDBC)、盒柱极差法(lacunarity of box column range,LBCR)、立方盒质量法(lacunarity of cube box mass, LCBM)和本文提出的盒柱平均值法(lacunarity of box column mean, LBCM)进行了对比分析。小样本正态分布拟合优度检验和小样本双边Student-t检验表明：由4种分维方法得到的分维值基本均呈正态分布,用4种分维方法得到的正常肝和肝癌超声图像分维的平均值均在0.05置信水平上差异显著；由5种孔隙度方法得到的最佳尺度孔隙度值,除LBCR方法得到的肝癌孔隙度值外,其他方法得到的孔隙度值均呈正态分布,但只有LCBM和本文提出的LBCM方法得到的正常肝和肝癌超声图像孔隙度平均值在0.05置信水平上差异显著。除FBM外,由其他分维方法得到的正常肝超声图像分维平均值均小于肝癌超声图像分维的平均值,同样,由LCBM和LBCM方法得到的正常肝超声图像孔隙度平均值亦小于肝癌超声图像孔隙度平均值。(2)SVM和ROC对分维和孔隙度的分类研究。为进一步确认上述4种分维方法和5种孔隙度方法表征肝脏超声图像纹理特征的能力,在此用受试者工作特征曲线(Receiver Operating Characteristic,ROC)和支持向量机(Support Vector Machine,SVM)对上述分维方法和孔隙度方法获得的正常肝超声图像和肝癌超声图像的分维值进行了ROC评估和SVM分类。评估和分类结果表明：分维方法中的FPS方法和孔隙度方法中的LBCM方法能获得较大的ROC曲线下面积(Area Under ROC Curve, AUC)和较高的分类准确率。(3)SVM和ROC对分形和孔隙度的组合因子的分类研究。用前述的4种分维值和LBCM孔隙度值作为单因子并由这4种分维值分别与LBCM孔隙度值构成组合因子(blanket+LBCM、FPS+LBCM、FBM+LBCM、 DBC+LBCM)作为分析因子,用SVM分类和ROC评估方法进行了对比分析,发现除FPS和LBCM单因子仍然得到较高的分类准确率和AUC外,组合因子FPS+LBCM和DBC+LBCM两种因子比其对应的单因子和其它组合因子有更高的分类准确率和AUC。(4)用ROC对三种多重分形方法进行对比研究。通过以盒合计值、小波系数和小波主要指数为测度对正常肝和肝癌超声图像进行多重分形分析,实验结果表明以盒合计值为测度矩阶g在[-1,1]范围内时获得了最大的AUC值,这表明盒合计值测度具有较强的捕捉肝脏超声图像纹理特征的能力,肝脏超声图像纹理特征属于低维的分形特征。更多还原

【Abstract】 Liver cancer is the main disease harming human heath and even imperiling human life, so that it is early accuracily diagnosed is urgent affairs. with the fast development of computer and image processing techniques, medical images have become the major information resources of liver disease dignosis and meanwhile it promotes the advance of model methods extracting image(including medical imagery) texture features. It will better make up the defect that the method of visually detecting liver ultrasound image does not far meet the requirements of accurate diagnosis. therefore, the research has become the hot issue.This dissertion deeply studied the classifation and recognition of liver ultrasonic images by means of computer-aided diagnosis techiniqus, taking normal liver and liver cancer ultrasonic images for samples and taking the fractal features of liver ultrasonic image texture for information. The research contents include:(1)Statistic analysis to fractal dimension lacunarity. the dissertion by comtrast analyzed4fractal dimension methods of good performance behaving in other fields about image texture feature exreaction-blanket, Fourier power spectrum(FPS), fractional Brownian motion(FBM), differential box counting(DBC). Meanwhile, it also analyzed another parameter in fractal geometry-lacunarity,5of its estimating methods-Mandelbrot, lacunarity of differential box counting(LDBC), lacunarity of box column range(LBCR), lacunarity of cube box mass(LCBM) and lacunarity of box column mean(LBCM). Small sample normal distribution goodness of fit test and small sample double sided Student-t test indicate that the fractal dimension values got by4methods all basically presented normal distributions, and the fractal dimension mean values of normal liver and liver cancer ultrasonic images attained with4fractal dimension methods all presented significant difference at0.05confidence level. Except for LBCR, the lacunarity values at best scale achieved with other lacunarity methods all indicated normal distributions, and yet the lacunarity mean values of normal liver and liver cancer ultrasonic image achieved with only LCBM and the proposed LBCM in this dissertion presented significant difference at0.05confidence level. Except FBM, the fractal dimension mean values of normal liver ultrasonic image all were less than those of liver cancer ultrasonic image attained with other fractal dimension methods, and likewise the lacunarity mean values of normal liver ultrasonic images were less than that of liver cancer ultrasonic images attained with LCBM and LBCM lacunarity methods.(2)Classification to fractal dimension and lacunarity with SVM and ROC. in order to verify the abilities of describing liver ultrasonic image texture features by4fractal dimension and5lacunarity methods related above, hereon one evaluated and classified the fractal dimension and lacunarity values of normal liver and liver cancer ultrasonic images achieved by above fractal dimension and lacunarity methods with Receiver Operating Characteristic(ROC) and Support Vector Machine(SVM). Evaluation and classification show that FPS of fractal dimension methods and LBCM of lacunarity methods achieved higher area under receiver operating characteristic curves(AUC) and higher classification accuracy.(3)The combined factors of fractal dimension and lacunarity with SVM and ROC. taking4fractal dimension values and LBCM lacunarity value for single factors and taking the combination of4fractal dimension factors with LBCM lacunarity respectively for the combined factors(blanket+LBCM、FPS+LBCM、FBM+LBCM、DBC+LBCM), one, by contrast, analyzed and evaluated the single factors and the combined factors with ROC and SVM. Evaluation and classification results indicate that except the single factors, FPS and LBCM, remain higher AUC and classification accuracy, the combined factors, FPS+LBCM and DBC+LBCM, hold higher AUC and classification accuracy than their corresponding single factors and other combined factors.(4)Contrastive analysis of3multifractals with ROC. the samples for normal liver and liver cancer ultrasound image were analyzed with multifractal method by taking box-sum, wavelet coefficient and wavelet leader for the measures. Experimental results indicate that the measure for box-sum attains maximal AUC value in the range of [-1,1] and hold the more strong ability catching liver ultrasound image texture feature as well as liver ultrasound image texture feature is fractal features of low dimension.31figures,23tables,127refences更多还原