【作者】 张红颖；

【导师】 孙济洲；

【作者基本信息】 天津大学 ， 计算机应用技术， 2007， 博士

【摘要】 图像配准是图像处理领域最重要、近年来发展最迅速的一项技术,在医学、军事、遥感、计算机视觉等众多领域得到了广泛应用。但是目前配准算法研究中存在很多亟待解决的问题,本文主要就其中制约医学图像配准的几个关键问题进行了深入研究。其中,针对基于互信息的多模态刚性配准算法中,插值计算使得目标函数变得不光滑从而使优化进程容易陷入局部极值、造成误配准的问题,本文提出了在多分辨率策略下使用遗传算法参数随多分辨率级数调整的改进优化方法。该方法可以有效地避免配准进程陷入局部极值、达到亚像素级精度。为了解决基于互信息的刚性配准算法容易陷入局部极值的问题,本文利用基于Shannon熵和Renyi熵的互信息各自作为配准相似测度所表现出的特点,提出一种基于混合互信息的配准方法。该方法在配准的不同阶段使用不同的互信息测度,并将微粒群优化算法与Powell优化相结合来对目标函数寻优。该方法可以快速、准确地得到配准变换。基于标记点的弹性配准存在标记点选取困难、变换模型不光滑准确等问题。为了解决这些问题,本文通过计算图像对应区域的相似性,来提取能够较精确地反映图像局部变形的对应标记点;并在此基础上,结合多层次B样条散乱数据插值,解决了配准精度和光滑性之间的平衡问题。利用互信息方法在多模态刚性配准中的成功应用,本文提出了一种有效地解决多模态弹性配准问题的改进Demons算法。该方法使用两幅图像间互信息对当前变换的梯度作为驱动图像变形的附加力,避免了Demons算法仅依靠图像灰度梯度变形、当梯度信息缺乏时图像变形方向不能确定的问题,从而得到更为精确的配准变换。在基于灰度的弹性配准算法中,针对高斯低通滤波对变形场进行正则化处理,在整幅图像上具有相同的正则化程度、不能反映图像真实变形的问题,本文提出了一种基于多层次B样条的正则化策略。该策略由粗到精地对所估计的偏移向量赋予不同的权值,从而局部地调节正则化程度,得到正确的配准变换函数。更多还原

【Abstract】 Image registration has been the most important and rapid developed technique in the field of image processing in recent years. Its application covers the areas of medicine, military affairs, remote sensing, computer vision, etc. However, there are still some difficult issues need to be studied further, especially for medical image registration, which are the main topics of this paper.Firstly, in multi-modality rigid registration algorithm based on mutual information(MI), interpolation operations result in non-smooth objective function, which makes optimization process get stuck into local extremes easily, thus wrong registration parameters are obtained. Aiming at this problem, a new registration algorithm is presented in this paper. The main idea of the new algorithm is to combine the genetic algorithm with multi-resolution strategy and the parameters of genetic algorithm are adapted along with the resolution level of images. The new algorithm can efficiently avoid registration process to get stuck into local extremes and obtain sub-voxel registration accuracy.Secondly, to solve the problem that MI-based rigid registration method gets stuck into local extremes easily, a new registration method based on mixed mutual information is proposed in this paper. This method uses different mutual information as similarity metric utilizing their characteristics of Renyi’s entropy and Shannon’s one. Moreover, the optimization method combining PSO with Powell is used to find the optimum. The proposed registration method can obtain more exact registration results than traditional mutual information based on Shannon’s entropy.Thirdly, for elastic registration problem based on landmark points, to obtain the corresponding landmark points presenting local geometric deformation of image precisely, a method is proposed in this paper. The method extracts landmark points through computing the regional similarity between two images. Furthermore, multilevel B-splines interpolation is applied to these landmark points to balance between the smoothness and accuracy of registration transformation.Fourthly, based on the successful application of maximization of mutual information in rigid multi-modality image registration, an improved“Demons”algorithm for elastic multi-modality images is proposed in this paper. The method adds additional external force defined as the gradient of mutual information between two images with respect to the deformation fields to drive the floating image to deform. In this way, the misregistration problem resulted by the original algorithm when transformation direction can not be determined due to the lack of intensity gradient information can be overcome.Fifthly, for intensity-based elastic registration problem, Gaussian smoothing is used to constrain the transformation to be smooth and thus preserve the topology of image. Aiming at the insufficiency of the uniform Gaussian filtering of the deformation fields, an automatic and accurate elastic image registration method based on B-splines approximation is proposed in this paper. In this approach, the regularization strategy is adopted by using multi-level B-splines approximation to regularize the displacement fields in a coarse-to-fine manner. Moreover, it assigns the different weights to the estimated displacements in terms of their reliability. In this way, the level of regularity can be adapted locally, so the estimated transformation is restricted to deformation satisfying the real-world property of matter.更多还原