作　　者： ; ; ; ;

机构地区： 华南理工大学自动化科学与工程学院

出　　处： 《计算机学报》 2019年第9期2087-2099,共13页

摘　　要： 患者随访CT图像之间的配准有助于提高诊断的可靠程度和改善治疗效果,是监测术后康复状况、确定或调整治疗方案等的前提.当前,先进的微分同胚形变配准算法的形变场驱动力仅依靠图像灰度和梯度信息,使得其对复杂形变的配准明显表现出驱动力不足、形变程度不高、鲁棒性较弱、配准精度低等问题.针对这些问题,该文提出将对数欧拉协方差矩阵(Log-Euclidean Covariance Matrices,LECM)描述符结合于配准模型的目标函数中构建了一种新配准算法模型,称为LECM Demons模型.该算法首先将图像每一像素的灰度信息、位置信息、一阶和二阶梯度范数信息映射为一个特征空间;然后,对特征空间采用积分图像方法快速计算每一像素的协方差矩阵;接着,通过矩阵的对数映射将协方差矩阵映射为Log-Euclidean空间中的特征描述符简称为LECM描述符,其对大的旋转、缩放、照度等变化具有不变性的特点,能够有效地描述图像的结构特性;最后,将待配准两图像对应的LECM描述符的欧氏距离作为一个新的匹配项添加到Log-Demons配准框架的目标函数,为大而复杂形变的图像配准提供了具有结构信息约束的形变场驱动力,同时确保了新目标能量函数的可微性.此外,为了进一步提高配准的收敛速度和精度,在新配准算法实现中采用了多分辨率优化策略.实验结果表明,LECM Demons算法对复杂形变具有较强的鲁棒性,配准精度与先进的形变配准算法相比平均改善50%以上,同时保持了较高的计算效率. The registration of CT images in patients is helpful to improve the reliability of diagnosis and to improve the therapeutic effect, which is the precondition of monitoring postoperative rehabilitation condition, confirming or adjusting treatment plan. At present, the deformation field driving force of the advanced deformation registration algorithm relies only on the image gray level and gradient information, which makes the registration of complex deformation appear to be lack of driving force, low deformation degree, weak robustness and low accuracy of registration. In order to solve these problems, a new registration algorithm, called LECM Demons Algorithm, is proposed to combine the Log-Euclidean Covariance Matrices (LECM) descriptor with the target function of the deformation registration model. Firstly, the gray information, position information, first-order and two-step-degree norm information of each pixel are mapped into a feature space;then the covariance matrix of each pixel is computed quickly by using the integral image method for the feature space. Secondly, the covariance matrix is mapped to the feature descriptor in the Log-Euclidean space by the logarithmic mapping of the matrix as the LECM descriptor, which is invariant to the large rotation, scaling and illuminance, and can describe the structure characteristic of the image effectively. Finally, the Euclidean distance of the LECM descriptor corresponding to the two images is added to the target function of the Log-demons registration frame as a new match. The image registration for large and complex deformation provides the driving force of deformation field with structural information constraints, while ensuring the differentiable of the new target energy function. In addition, in order to further improve the convergence speed and accuracy of registration, the proposed algorithm first uses the rigid registration algorithm based on the pixel gray level to make the global rough registration. The result of rough registration is used as the initia

关 键 词： 医学图像配准 协方差矩阵 多分辨率策略 鲁棒性

领　　域： []