基于遗传算法及最大互信息的医学图像配准研究

【作者】 杨日容；

【导师】 杨承志；

【作者基本信息】 昆明理工大学 ， 控制理论与控制工程， 2005， 硕士

【摘要】 在临床诊断治疗中,往往要求对病人的病变部位进行多次成像,以获取互补、有效、全面的信息,提高医生的诊断治疗效果,这就需要进行医学图像的信息融合,把多幅图像的信息融合在一起,并在一幅图像上体现多方面的信息。作为医学图像信息融合的基础,医学图像配准 (medical image registration) 具有重要的临床应用价值,不仅可以用于诊断治疗,还可以用于病理变化的跟踪和治疗效果的评价等多方面。医学图像配准已经成为当前医学图像处理领域中的研究热点,它主要是寻找两幅图像之间的一个最优变换,使得其中一幅图像在这种变换下和另一幅图像达到空间位置和解剖结构的一致,实质上这是一个最优化问题。医学图像配准的研究具有一定的复杂性和困难度,虽然目前已经提出了许多的算法,但每种配准方法都是针对某一特定问题而设计,具有一定的局限性,而且在配准速度、配准精度等方面都不能同时达到理想的效果。因此,针对医学图像配准的研究是一项极富挑战性的课题,许多研究者都在该领域不断地进行探索。 特征空间、几何变换、相似性测度以及优化算法是设计医学图象配准方法时要考虑的四个主要因素。已经提出的医学图像配准方法主要可以分为两类:基于特征的配准和基于灰度统计的配准,其中用得较多的是基于特征的配准。基于特征的配准方法简单且易于实现,但配准精度受到特征提取的分割步骤的影响。本论文引入一种最新的基于灰度统计的图像配准模型:互信息配准模型,基于该模型的配准方法称为最大互信息法,它利用信息论中的互信息作为两幅配准图像之间的相似性测度,不需要对图像进行分割等预处理,几乎适用于任何不同模式图像的配准,能够得到很好的配准效果。最大互信息法的基本思想是:当基于共同解剖结构的两幅图像达到最佳配准时,互信息达到最大。本论文给出了最大互信息法的相关概念和理论,并分析了采样率对互信息的影响以及互信息和图像重叠度之间的正比关系。在图像配准过程中,采用了三线性PV插值来避免图像变换所带来的灰度直方图计算困难的问题。 医学图像配准中采用的优化算法很少涉及到全局优化算法,主要以局部优化算法居多,但这些局部优化算法容易陷入局部极值,从而导致图像的误配准。本论文采用的优化算法是具有良好全局搜索能力的遗传算法。但是,遗传算法由于其存在容易早熟、收敛速度慢的不足,在医学图像配准的应用中有一定的缺陷。本论文通过在编码、遗传操作算子等方面对标准遗传算法进行改进,提高了算法的性能。 本论文针对人脑的二维图像设计了一种遗传算法和最大互信息相结合的医学图像配更多还原

【Abstract】 In clinical diagnosis and treatment, it is usual to have more images of focus of patient. In order to obtain the complementary and effective and comprehensive information, it is necessary for more medical images to fuse information. As the basis of images information fusing, medical images registration have important applied values, it is a hot point in the medical images treatment. Medical images registration want to look for the best transformation parameters, and match corresponding points of images, and express many-sided information in one image. Medical images have complexity and difficulty. Although many methods have been brought forward, they have some limitation because every registration method is only designed for the given problem. Moreover, registration precision and speed can not obtain perfect effect at the same time. Therefore, it is a challenging task to study medical images registration. Many investigator explore in the domain.Medical images registration methods have four main factors: features space, geometry transformation, similarity, and optimizing algorithm. Now, Many registration methods which have been brought forward can be divided into two kinds: the registration methods base on image features and the registration methods based on gray statistics. And the former is often used. The kind of registration method is simple and can be arrived easily, but the registration precision is limit to the segmentation when the image features are acquired. The paper bring into the newest registration model based on gray statistics: mutual information registration model, and the method which based on it is called the maximal mutual information method. The method takes mutual information in the information theory as similarity between two registration images, and it does not need processing ahead of time such as segment, and it almost suits to any kinds of image registration. The primary ideas of the method is that mutual information will achieve the max value when the two images based on the mutual anatomy structure. The paper gives the concepts and theories about mutual information, and thorough analyzes the effect sampling rate of on mutual information and the ratio between mutual information and images overlaying. Moreover, Trilinear Partial Volume Distribution Interpolation is took in order to avoid the difficulties of calculating gray because of image transformation.Medical images registrations usually use local optimizing algorithms, and the global optimizing algorithms are less concerned. But these local algorithms are easy to fall into thelocal supreme value, so it will lead to the wrong registration. The paper uses genetic algorithm which has good global searching ability. Because of the earliness , slow searching speed and long running time, genetic algorithm has some defects when it is applied in medical images registration. The paper improves the standard genetic algorithm from coding, genetic operating ,and so on. It improves the capability of the algorithm.The paper designs a registration algorithm which links improved genetic algorithm and the maximal mutual information method for two 2D brain images. It takes mutual information model, and bases on the intensities of the images, and uses improved genetic algorithm to search the best transformation parameters. It also uses the maximal mutual information as the aim function to guide searching the best transformation parameters, and the effectiveness of the algorithm is proved. The paper also has a useful try for applying the global optimizing algorithms in the medical images registration.更多还原