一种面向医学图像非刚性配准的多维特征度量方法

医学图像的非刚性配准对于临床的精确诊疗具有重要意义.待配准图像对中目标的大形变和灰度分布呈各向异性给非刚性配准带来困难.本文针对这个问题，提出基于多维特征的联合Renyi α-entropy度量结合全局和局部特征的非刚性配准算法.首先，采用最小距离树构造联合Renyi α-entropy，建立多维特征度量新方法.然后，演绎出新度量准则相对于形变模型参数的梯度解析表达式，采用随机梯度下降法进行参数寻优.最终，将图像的Canny特征和梯度方向特征融入新度量中，实现全局和局部特征相结合的非刚性配准.通过在36对宫颈磁共振（Magnetic resonance，MR）图像上的实验，该方法的配准精度相比较于传统互信息法和互相关系数法有明显提高.这也表明，这种度量新方法能克服因图像局部灰度分布不一致造成的影响，一定程度地减少误匹配，为临床的精确诊疗提供科学依据.

A Metric Method Using Multidimensional Features for Nonrigid Registration of Medical Images

Nonrigid registration of medical images has great significance for accurate diagnosis and therapy in clinic. It is difficult to register the images containing large deformation of object region and data anisotropy. According to this problem, an algorithm of nonrigid registration based on joint Renyi α-entropy is proposed in this paper, which combines global features with local features. Firstly, minimum spanning tree is employed for construction of joint Renyi α-entropy. A new metric is built on multidimensional features. And then, the analytical derivative of the new metric with respect to the parameters of deformation model is derived, in order to find the optima by a stochastic gradient descent method. Finally, Canny feature and gradient orientation feature of images are merged into the new metric, which implements nonrigid registration including global and local features. Experiments are performed on 36 cervical magnetic resonance (MR) image pairs. Compared to the traditional mutual information and correlation coefficient, the registration accuracy is improved significantly. It also manifests that the proposed method is able to overcome the adverse effects of local intensity inhomogeneity, and provides scientific evidence for accurate diagnosis and therapy in clinic, due to reducing mismatch in some degree.