Abstract: 3-D medical image registration based on mutual information is researched in this paper and local maxima are overcome. The MI’s formula is improved to reduce computation cost, and the replacement of Powell’s search direction is also rectified to reserve the search direction and escape local maximum, initial set of rotation can avoid local maximum induced by interpolation anti-facts. Threshold is set to reduce iteration region, morphology method is used to reduce noise of PET image, muti-resolution method is adopted to accelerate the registration. The results show that the modified MI measure can reach sub-voxel precision, and the registration time is reduced obviously, is a robust and precise method.

Key words: Medical image registration, Normalized mutual information, Morphology, Powell search algorithm, Interpolation anti-facts

摘要： 本文研究了基于互信息的三维医学图像配准，并克服了的局部极值问题。本文改进了互信息和归一化互信息的公式，减小了计算量；对Powell算法的方向替换策略进行矫正，最大限度的保持原搜索方向；采用预设旋转量的方法有效解决了插值赝像局部极值问题。采用背景阈值策略，减小了互信息的计算区域，应用形态学方法，去除了PET图像的背景伪迹，使用了多分辨率策略，有效地提高了配准的速度。实验表明，用本文改进的算法进行三维医学图像配准可以达到亚像素精度，且在速度上有了明显的提高。

关键词: 医学图像配准, 归一化互信息, 形态学, Powell算法, 插值赝像