一种新的结构正割法在时间序列图像配准中应用

Application of new totally structured secant method to registration of time series images

脑功能成像是进行神经科学研究的一项重要技术.在脑功能图像的数据分析中,时间序列图像的配准精度是对脑功能图像进行统计分析成功与否的关键.针对一种典型的非线性最小二乘形式的配准测度,引进一种新的结构正割法来计算其极小值.该方法给出了目标函数海森阵二阶信息项的一种新的较佳近似,而并非像高斯-牛顿法那样将其直接舍去,从而提高了计算精度;并且在近似的过程中,校正矩阵始终保持对称正定,所以在搜索下降方向的时候,总会找到一个解.模拟计算结果表明该方法配准精度高,计算时间较短,可以较好地解决图像配准问题.

Brain functional imaging is an important tool in the study of nerval science, and it is necessary to register time series images. The new totally structured secant method is adopted to optimize the nonlinear least squares registration criteria, which gets a better approximation to the Hessian of the criteria by secant update, while the conventional method--Gauss-Newton algorithm directly gets rid of the second-order information of the Hessian. At the same time, the update matrix is always symmetrical and positive which makes it reversible, so a descent direction can be found in each step. The simulation registration results show that the algorithm can quickly obtain the best registration parameters with high precision.