三维旋转估计及其在医学影像配准中的应用  

Three dimensional rotation estimation and its application in medical image registration

作　　者：张文妙韵 周武[2] 戴培山[1] 

机构地区：[1]中南大学地球科学与信息物理学院,湖南长沙410083 [2]中国科学院深圳先进技术研究院,广东深圳518055

出　　处：《中国医学物理学杂志》2015年第4期542-549,共8页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(61302171;81171420);中国博士后科学基金(2013M540740);自然科学基金广东省联合重点资助项目(U1301258)

摘　　要：目的:针对基于迭代优化的影像配准方法在迭代初始值离目标值较远的情况下普遍需要消耗较多的时间或者易陷入局部极值的问题,提出了一种基于局部梯度信息统计相关的三维旋转估计方法。方法:首先通过构建三维高斯梯度滤波器求出影像的局部梯度,然后根据局部梯度计算出影像的方向分布直方图,最后由两幅影像对应的方向分布直方图相关匹配的结果估计旋转差异。将该方法应用于基于优化的医学影像刚体配准中,比较在没有初始值估计和由该方法提供初始值估计的两种不同情况下配准迭代所需的时间以及配准后的效果。结果:经过测试,当虚拟旋转的角度不是很大时(比如在20°以内),该方法估计的结果与虚拟变换设定值之间的偏差通常不会超过3°,一般在2°以内。同时,该方法在医学影像配准中的应用明显减少了迭代所需的时间,且相比较由于不了解迭代初始值的信息而将其假设为零的情况,通过应用该方法来估计初始值有效降低了优化在迭代过程中陷入局部极值的可能。结论:当影像之间的旋转差异不是太大时(比如不超过20°),可以采用本文提出的方法来估计旋转,并且能获得较为理想的估计结果。此外,该方法在医学影像刚体配准中的应用提高了迭代优化的效率,也有助于进一步避免局部最优的问题。Objective To propose a three dimensional （3D） rotation estimation method based on statistical correlation of local gradient information, because the image registration based on iterative optimization generally consumes much time or easily fall into the local extrema when the iterative initial values are far away from the target values. Methods A 3D Gaussian gradient filter was constructed to obtain the local gradients of images. And the orientation distribution histograms were calculated based on the obtained local gradients. Finally, the rational differences were estimated by the relevant matching results of the corresponding orientation distribution histograms of these two images. This method was applied into rigid medical image registration based on iterative optimization to provide estimated initial values, and then the time needed for registration iteration and the registration effects under different condition with or without estimated initial values were compared. Results When the angle of virtual rotation between the two images was not too large, such as within 20°, the deviation between the results estimated by this method and virtual transformation set value was less than 3°, and it was always in the range of 2°. Furthermore, this method obviously reduced the time needed for the iteration in the medical image registration, and effectively reduced the possibility of falling into the local extrema during the iteration, compared with the situation of just assuming iterative initial values to be zero. Conclusion When the rotation angle between the two images is not too large, such as within 20°, the proposed method can obtain satisfactory estimated results and its application in rigid medical image registration improves the efficiency of iterative optimization and reduces the appearance of local extrema.

关 键 词：三维旋转估计 梯度方向 直方图匹配 医学影像配准 迭代优化 

分 类 号：R312[医药卫生—基础医学] TP391.41[自动化与计算机技术—计算机应用技术]