【作者】 杨振森；

【导师】 周康源； 冯焕清；

【作者基本信息】 中国科学技术大学 ， 生物医学工程， 2009， 博士

【摘要】 前列腺癌是一种中老年男性的常见疾病,并且只有在癌症初期诊断出来才能够有效提高患者的生存率,通常使用直肠超声穿刺技术来进行前列腺疾病的早期检查。大量的研究表明,使用计算机辅助诊断技术(CAD)可以进一步提高病理诊断的准确性。前列腺病变CAD系统通常使用病变区域的形态和灰度纹理统计特征来对病变区域进行分类判别。由于超声图像对比度较低,仅使用这些特征很难有效地区分病变区域。因此本文研究超声图像的病变区域在不同尺度下和不同频率域中所具有的不同特征,并将它们有效地用于前列腺病变的辅助诊断。本文所完成的主要研究和创新点描述如下:(1)对直肠超声(TRUS)图像进行预处理。针对超声图像中斑点噪声较为严重的特点,本文提出了一种基于改进的各向异性扩散方程的自适应斑点抑制算法,通过使用区域中值代替原像素点和加入对角梯度算子,有效的抑制了斑点噪声,同时也增强了细节信息;加入了保真项,使得算法在有效地抑制斑点噪声的同时4还具有良好的鲁棒性,同时能够有效保持边界几何信息,解决了传统算法受迭代次数影响较大的问题。(2)对前列腺超声图像进行分割,以得到前列腺的边界轮廓。首先提出了一种基于改进测地线模型的全自动前列腺边界轮廓分割算法,将径向浅浮槽算法和满水法填充算法结合获得超声图像中前列腺的大致目标轮廓,来作为主动轮廓模型的初始水平集,并且构造基于区域信息的符号压力函数来代替边界停止函数,同时加入了基于边界梯度信息的能量项,能够有效地克服弱边界的问题。该算法用二值水平集实现,使得算法的稳定性大为提高,并且计算量大大降低。针对对比度较差的超声图像,又提出了一种基于改进测地线主动轮廓模型的半自动分割算法,该算法使用手工勾画的粗略边界作为初始水平集,使用变分法,消除了测地线模型中的重初始化过程,使得算法运算速度有了较大的提高,根据区域先验知识改进了测地线的边界停止函数,能够有效避免边界泄漏问题,并使用最小方差来提高分割算法的精度。(3)对分割后的图像进行前列腺病变特征参数的提取。不同于以往的研究仅仅使用灰度共生矩阵(GLCM)特征或者灰度级差矢量(GLDV)特征来对病变进行识别,本文所提取的特征参数包括小波域统计特征参数、各种尺度频率域特征参数、梯度图像特征参数等。由于特征参数之间存在一定的冗余性,为了有效地分析和利用所提取出的特征,使用了特征维度约减算法来选择最优的特征参数子集。(4)根据所提取的特征参数来对前列腺病变区域的良恶性进行分类判断。使用SVM算法和AdaBoost算法对所提取出来的病变区域特征参数进行分类,并分析对比了不同特征参数对分类器判别结果的影响和不同分类器之间的分类性能,以寻找最优的组合。按照上述四个步骤,本文对255幅前列腺直肠超声图像(其中125幅为恶性病变图像,130幅为良性病变图像)进行实验。实验结果表明,联合使用本文所提取出来的特征参数集的辅助诊断系统,性能明显优于只使用灰度级差矢量特征参数的系统,这表明本文所提出的基于直肠超声图像的系统对前列腺病变区域的良恶性具有较好的区分能力,能够为医生的临床诊断提供辅助依据。更多还原

【Abstract】 Prostate cancer is one of the common diagnosed malignancies in middle-aged and elder men,and the survival rate of the patients can only be enchanced by detection in the early stage of cancer.Nowadays,TransRectal Ultrasound(TRUS) puncture technique is one of the most frequently used methods for the early detection of prostate cancer.Many researches show that the accuracy of the pathology detection can be further enhanced by using computer aided diagnosis(CAD) technology.The prostate cancer CAD system usually performs the diagnosis based on the morphology and grayscale texture statistical features of the pathological areas.However,the CAD system only utilizes the aforementioned features can’t differentiate the pathology areas effectively because of the poor contrast in the TRUS images.Therefore,this dissertation studies the different characteristics of the pathological areas in the different scales and frequency domains to assist the diagnosis of prostate cancers effectively.The main research work and contributions of this dissertation can be summarized as follows:(1) Preprocessing the TRUS images.Aiming at the characteristics that the ultrasound images have a lot of speckle noises,an improved anisotropic diffusion adaptive speckle reduction algorithm was proposed.The algorithm used the medians of regions to replace the original pixels and added the diagonal gradient term,which would suppress the speckle noises effectively and enhance the detail information.The algorithm combined a fidelity term,which made the algorithm suppress the speckle noise effectively and more robust,moreover,the algorithm could maintain the boundary geometry information effectively and solve the problem that the results of the conventional algorithms were affected by the iterative numbers greatly.(2) Segmenting the TRUS images to get the prostate boundary.This dissertation proposed an automatic prostate boundary segmentation algorithm based on the improved geodesic active contour.Firstly,a coarse contour could be gotten by combining radial bas-relief and region fill algorithm to the TRUS images,the contour could be used as the initial contour for the active contour model.Then the problem of weak edges could also be efficiently solved by presenting a new region-based signed pressure forces function to replace the edge stopping function and incorporating an energy term based on boundary gradient information.The algorithm was implemented by binary level set function,which reduced the expensive computational cost of re-initialization of the conventional level set and enhanced the stability of the algorithm.A semi-automatic segmentation algorithm based on improved geodesic active contour model was also proposed to deal with the TRUS image with poor contrast.The algorithm used the manual-drawing boundary as the initial level set.The need of the costly re-initialization procedure is completely eliminated by using variational formation,thus increased the speed of the evolvement.The edge stopping function in the geodesic active contour model was revised by incorporating a priori area information to avoid the problem of edge leaking.The accuracy of the algorithm was improved by the minimal variance term.(3) Extracting the characteristic parameters of the prostate pathology from the segmented images.Other than the past studies only extracted the gray level coocurrence matrix(GLCM) and gray level difference vector(GLDV) features,the dissertation extracted parameters including wavelet statistical characteristics, characteristics from all scale-frequency domains,and characteristics from gradient images.There existed the redundancy among the extracted feature set,so the study used feature dimension reduction method to select the optimal feature subset for the sake of utilizing the extracted feature sets efficiently.(4) Utilizing the feature subset extracted in the previous step to aid the classification of prostate pathological areas.The dissertation used SVM and AdaBoost to classify the extracted characteristics parameters of pathological areas.The study compared the affect of different characteristic parameters to the results of the classifier and the classification performance between the different classifiers,through which to look for the optimal combination of characteristic parameters and classifier.Based on the aforementioned four steps,255 prostate TRUS images including 125 benign pathological areas and 130 malignant ones were studied in the dissertation, the experiment results showed that the performance of CAD system which used the characteristic parameters including wavelet statistical characteristics,characteristics from all scale-frequency domains,and characteristics from gradient images was much better than the system which only used the GLCM or GLDV characteristics. Therefore,this indicated that the system proposed by the dissertation performed well in the classification of malignant and benign between the prostate pathological areas, and it could be used as a supplementary basis for the clinical diagnosis of the prostate cancers.更多还原