基于3D-PCNN和互信息的3D-3D医学图像配准方法

摘要

基于特征的配准方法配准精度低,基于互信息的配准方法虽然配准精度高但计算量大且易陷入局部极值.针对此问题,提出一种基于脉冲耦合神经网络(PCNN)和互信息的由粗到细的3D-3D医学图像配准方法.首先,将2D-PCNN模型扩展成能直接处理三维图像的3D-PCNN模型.然后,采用Eckhom简化输入部分方式对扩展的3D-PCNN模型进行简化,并用线性衰减阈值代替指数衰减阈值,降低了PCNN网络计算复杂度.为了自适应确定参数值,从待处理的三维图像的二维切片图像中随机选择一幅切片图利用二维参数优化方法求出2D-PCNN参数值,并将此参数值作为3D-PCNN的参数值.最后,后利用PCNN网络点火集群的平移、旋转、尺度缩放、扭曲等不变特性将各次迭代点火集群的几何重心作为特征点,实现图像粗配准,获得初始配准参数,将此粗配准参数结果作为基于互信息配准搜索算法的初始参数值,使得搜索始终围绕全局最优值附近进行,进一步微调细化参数,得到最终较高精度的配准结果.%The registration method based on characteristics is of low accuracy,however the mutual information-based method is of high accuracy,but the computed quantity is large and it is easy to get into local maximum.In order to solve this problem,a coarse-to-fine 3D-3D medical image registration method based on 3D-PCNN (Pulse-Coupled Neural Network) and mutual information was proposed.Firstly,the 2D-PCNN model was extended to 3D-PCNN model which could directly deal with three-dimensional images.Secondly,considering the complexity of 3D image processing,the Eckhorn's method was used to simplify the input parts of the extensional 3D-PCNN model.Besides,the linear attenuation threshold was adopted instead of exponential attenuation threshold,which reduced the computing complexity.To adaptively determine 3D-PCNN's parameters,one slice image was randomly selected from 2D slices of the 3D image,and 2D-PCNN's parameters optimization method was used to determine the values of 3D-PCNN's parameters.At last,based on PCNN's clusters with the invariant characteristics of translation,rotation and distortion,the geometric centers of gravities of PCNN's clusters in each iteration were used as the feature points to achieve image registration.The coarse results provide a near-optimal initial solution for fine-tuning process using mutual information.According to further fine-tuning refining parameters,the final registration results with higher accuracy were gained.