融合SIFT特征的熵图估计医学图像非刚性配准  

Entropic graph estimation integrated with SIFT features for medical image non-rigid registration

作　　者：张少敏[1,2,3] 支力佳[1,2,3] 赵大哲[1,2,3] 林树宽[1] 赵宏 

机构地区：[1]东北大学信息科学与工程学院,沈阳110004 [2]医学影像计算教育部重点实验室(东北大学),沈阳110004 [3]国家数字化医学影像设备工程技术研究中心,沈阳110004

出　　处：《中国图象图形学报》2012年第3期412-418,共7页Journal of Image and Graphics

基　　金：国家自然科学基金项目(60873009)

摘　　要：配准准确性是医学图像配准算法的一项重要指标,像素灰度是目前图像配准中广泛使用的特征,但是灰度特征来源单一,而且忽略空间信息,在一些情况下容易产生误配。针对这个问题,本文提出一种融合SIFT特征的熵图估计医学图像非刚性配准算法。该算法首先使用基于互信息的刚性配准算法对两幅待配准图像进行粗配;然后,在采样点上提取像素灰度和SIFT高维特征,并在此基础上构造k-最邻近图(kNNG);最后,使用k-最邻近图来估计α互信息(αMI)。实验结果表明:和传统的基于互信息和像素灰度的刚性配准算法,基于熵图估计和单一像素灰度特征的非刚性配准算法相比,本文提出的算法具有更高的配准准确性。Accuracy is important for the regrstration of medical images. Pixel gray values are a widely used feature in image registration. However, the gray values come from a single source and ignore the spatial information. In some cases, it will cause misalignment. To solve the problem, entropic graph estimation integrated with SIFT features is proposed as a medical image non-rigid registration algorithm. In the algorithm, mutual information based rigid registration is used to roughly register two images. Then the pixel gray value and the SIFT features are extracted to form a k-nearest neighbor graph （kNNG）, which is used to estimate a-mutual information （aMI）. Comparison results of the images obtained from lung CT images and brain MRI images showed that the proposed algorithm provides better accuracy than both, the conventional rigid registration algorithm based on mutual information and the non-rigid registration algorithm based on entropic graph estimation and single pixel gray values.

关 键 词：医学图像配准 SIFT描述子 k-最邻近图 α互信息 

分 类 号：TF391[冶金工程—冶金机械及自动化]