【作者】 张海哲；

【导师】 于明；

【作者基本信息】 河北工业大学 ， 微电子学与固体电子学， 2004， 硕士

【摘要】 随着医学影像技术的快速发展，出现了多种模态的医学影像。医学上，急需把这些描述不同解剖或功能信息的图像快速准确的融合起来，为临床诊断和手术治疗提供更加全面准确的信息。医学图像配准是将两幅图像的对应特征点达到几何对齐，它是融合的关键部分。 本文介绍了配准中常用的几何变换形式、优化搜索方法、各种相似性测度、配准的一般分类原则及国内外面临的问题和发展方向。然后对现有的配准算法及相关技术进行了整理，按是否提取图像特征为依据将配准方法分为基于图像特征的配准和基于体素的配准。 在基于体素的方法中，最大化互信息法是目前研究的热点，本文对最大互信息法的性能做了详细的分析。在添加了不同噪声干扰时，它表现出了很好的抗噪性；对于不同分辨率的图像配准，分辨率越高，配准精度越高；PV插值技术导致互信息相邻整数变换之间的非单调性。由于最大互信息方法并没有充分地考虑图像的空间信息，本文研究了其它的基于互信息的方法：互信息求导法、高阶互信息法等。 最大互信息法计算的是图像的重叠区域。当重叠区域随着视场的改变发生变化时，互信息各项的值也都发生变换，最大互信息方法没有充分考虑各项变化之间的关系，出现了误配准的情况。针对此问题，本文采用了一种基于归一化互信息配准的新方法。通过与其它几种相似性测度的比较，该方法对图像重叠区域的变化具有不变性，从而使配准更具鲁棒性。本文在理论和试验分析的基础上得出结论该方法优于最大互信息法。更多还原

【Abstract】 With the rapid development of imaging technology, there are more and more different modality medical images in application fields. Integrating those images is helpful to improve the accuracy of clinical diagnoses and surgical therapies .Image registration is the key part of integration. It’s a geometrical transformation aligning two images according to the corresponding features.This thesis introduces several geometrical transformations, optimization methods, similarity measures. According to whether extracting images’ feature or not , we divide present methods into two parts, feature extraction based (FEB) methods and voxel similarity based (VSB) methods.Maximum mutual information(MMI) is the most popular method in VSB. In this thesis we analyse its characters. It shows high performance in case of adding different noise to the floating image. When two images have different resolutions, the higher resolution the images have, the better result it turns out ,but it will consumes much more time. We also focus on partial volume (PV) interpolation algorithm .The value of mutual information(MI) doesn’t increase or decline in one direction at translation from one integer to the nearest integer. MMI doesn’t consider the spatial information, we introduce two other MI based methods: derivation of MI and high-order MI.MMI computes the overlap area of two images. The changing of the overlap with different fields of view leads to the value of each entry in MI changed. MMI doesn’t consider the relations about these changes. It shows misalignment. In this thesis we analyse this situation. Then we normalized the mutual information, which testified an overlap invariant measure, so we use normalized mutual information(NMI) as a new similarity measure. After theoretical and experimental analyses we draw a conclusion it’s a better method than MMI in multi-modality image registration.更多还原