胰腺超声内镜计算机辅助诊断系统研究

【作者】 蔡哲元；

【导师】 余建国；

【作者基本信息】 复旦大学 ， 医学电子学， 2010， 硕士

【摘要】 胰腺癌是严重危及人类健康的重大疾患,其早期发现、早期诊断、是长期困扰医学界的难题。目前的诊断方法中,超声内镜检查术对胰腺癌具有较大的诊断价值,但基于超声内镜图像的诊断受医生经验和主观因素影响大,而经穿刺的细胞学检查有一定创伤性。本文研究并初步开发了胰腺超声内镜计算机辅助诊断系统,旨在创建各种客观、量化的诊断指标以及正确的描述解释超声内镜图像的方法,提高胰腺癌超声内镜早期诊断的准确性。本文的研究内容主要由四部分构成：胰腺超声内镜的纹理特征提取、特征选择、分类评价系统和软件系统开发,并对各部分存在的一些问题进行研究和改进。在纹理特征提取部分,将各类文献中广泛使用的用于纹理分割与纹理分类的特征引入胰腺超声内镜图像分类,提取9大类74种纹理特征,并取得了较好的分类结果。之后对传统分形特征进行研究与改进。对Lee提出的基于M带小波变换的分形特征进行改进,引入多重分形维数并进行特征筛选,提出了基于M带小波变换的多重分形特征。基于本研究分形特征矢量的分类,在运行时间和分类准确率上均优于基于传统分形特征的分类。在特征选择部分,针对特征选择中传统可分性判据对线性不可分数据集评价能力差的特点,提出了基于核空间距离测度的可分性判据。在核空间中计算两类样本点之间的距离,并以距离的大小评价子集的分类性能。本文的方法体现了基于核方法的优越性,尤其提高了在小样本与线性不可分数据集上的特征选择能力,同时由于核空间距离测度随核参数单调变化,便于消除核参数的影响,相比Wang提出的核散布矩阵测度,大幅度降低了运行时间,同时保留了核方法的优越性。根据医生在临床上对分类评价的需求,提出了基于广义隶属度函数的模糊模式识别方法实现胰腺超声内镜分类评价系统。通过特征模糊化提高鲁棒性,使用训练样本构造合理的目标函数和惩罚项,通过多参数优化的方法得到模式广义隶属度函数的参数,并通过评价函数对样本进行分类评价。在以上算法研究的基础上,编程实现计算机辅助诊断软件系统开发。使用标准C++语言在Visual Studio环境下实现本文提出的特征提取与分类算法。使用MFC、GDI+、ODBC等技术实现浏览,计算机辅助诊断,放大,测量和数据库连接等多种功能模式,方便医生临床使用。更多还原

【Abstract】 Pancreatic cancer is a major disease which seriously threatens human health. The early detection and diagnosis are problems which have long plagued the medical profession. In current diagnostic methods, the endoscopic ultrasonography(EUS) have greater diagnostic value over pancreatic cancer. However, the diagnosis based on EUS images is affected considerably by the doctors’ experience and subjective factors, and the cytology examination is invasive. In this paper, we focus on the research and basic development of the computer aided diagnosis(CAD) system of pancreatic EUS images, in order to create a variety of objective, quantitative diagnostic indices and an appropriate method to describe and understand the EUS images, and finally, increasing the accuracy of early diagnosis of pancreatic cancer through EUS.The proposed methods are composed of four parts:textural feature extraction of the pancreatic EUS images, feature selection, classification and evaluation system and the software development. Study and improvement on each part have been done for specific problems.As for the textural feature extraction, textural features, which have been widely used in different papers for textural segmentation and classification, are introduced into the area of pancreatic EUS image classification. 74 texture features of 9 categories are extracted, and good classification results are achieved using these features. Further work has been done on the study and improvement of tradition fractal features. By modifying the M-band wavelet transform fractal feature based on the fractal dimension(Lee,2003), the multi-fractal dimension was presented with the feature selection to obtain the multi-fractal feature vector of M-band wavelet transform. Experimental results showed that the classification based on the proposed fractal feature outperformed those based on the traditional fractal feature in both executing time and classifying accuracy.In the part of feature selection, the kernel distance measure is proposed as a new type of class separability, considering the poor performance of traditional class separability on linearly non-separable data set. The distance of samples from two classes is measured in the kernel space, and used to evaluate the separability of subsets. The proposed method embodies the advantages of kernel-based methods, especially on small sample set and linearly non-separable data set. Meanwhile, the kernel distance measure changes with the kernel parameter monotonously, so that it is 4 easy to eliminate the influence of kernel parameter. Compared with kernel scatter matrix measure(Wang,2008), the proposed method is much faster in running time and retains the advantages of kernel-based methods as well.According to the clinical requirements of classification and evaluation system, the fuzzy pattern recognition based on generalized membership function is proposed to realize the classification and evaluation system of pancreatic EUS images. The features are made fuzzy to increase the robustness. The training samples are used to construct a reasonable objective function and item of punishment. The unknown parameters of generalized membership function is estimated through optimization, and finally the classification and evaluation is implemented by the evaluation function.Based on the above research on algorithm, the Computer Aided Diagnosis Software System is developed. The proposed feature extraction and classification algorithm is programmed using standard C++ language in the environment of Visual Studio. The different function mode of browsing, CAD, zooming, measuring and database connecting is realized through the technology of MFC, GDI+, ODBC, etc, to facilitate-medical clinical use.更多还原