在医学多图谱配准中, 为了改善因初始位置差异较大、形状复杂和局部残缺导致的配准效率低和精度差的问题, 本文采用了先粗配准再精配准的处理策略, 在主成分分析法(PCA)实现粗配准的基础上, 提出了基于双向距离比例的迭代最近点(ICP)的精配准算法。精配准算法中, 首先采用KD-tree进行最近邻搜索以提高对应点对的搜索速度, 然后为每个点提出了双向匹配方法并计算其双向距离和比值, 为进一步提高配准精度, 引入了一个指数函数判断点对正确匹配概率, 最后运用奇异值分解法(SVD)计算最终变换矩阵。为了验证算法的可行性和有效性, 分别设计了不同缺损程度的斯坦福点云数据实验和两组CT心脏点云数据配准实验, 结果表明本文方法较经典ICP算法的平均误差减少约21%, 较TrICP算法减少约13%, 在心脏点云数据配准实验中, 本文方法较TrICP算法的15.5 s加快到1.77 s。因此本文方法在解决三维心脏点云数据的配准问题中具有良好的效率、精度和稳定性。

In medical multi-atlas registration, to improve the limitations of low efficiency and poor accuracy caused by large initial position differences, complex shapes, and local residual differences, a fine registration algorithm that used Iterative Closest Point (ICP) was proposed based on the bidirectional distance ratio. The proposed algorithm was based on the coarse registration method followed by fine registration, where the former was processed by principal registration analysis.In the fine registration algorithm, the K-Dimensional tree was initially used to perform a nearest-neighbor search to improve the searching speed of corresponding point pairs. A bidirectional matching method was then proposed for each point, and the bidirectional distance and ratio were calculated. To further improve the accuracy of the registration, an exponential function was introduced to determine the probability that the point pair belongs to the correct match. The final transformation matrix was then obtained using Singular Value Decomposition. To evaluate the feasibility and effectiveness of the algorithm, experiments were designed using Stanford point cloud data and two sets of CT cardiac point cloud data registration. The results show that the average error during registration is reduced by 21% using this method compared to the classical ICP algorithm, which is 13% lower than the error obtained using the trimmed ICP (TrICP)algorithm. In the cardiac point cloud data registration experiment, this method is accelerated to 1.77 s compared to the TrICP algorithm, which has a value of 15.5 s.Therefore, the proposed method has high efficiency, accuracy, and stability in solving the registration problem associated with the three-dimensional cardiac point cloud data.