Abstract:

Aiming at the problems of slow convergence and easy to fall into local maximum in the traditional registration method for 3D multi-modal image, a method based on Fully Convolutional Networks（FCN） and mutual information algorithm is proposed. The FCN model is used to extract the deep features of 2D images and perform coarse registration. The registration result is used as the initial search point of the mutual information algorithm, which provides a near-optimal initial solution. The mutual information algorithm is used to further fine-tune the parameters to obtain the best 3D registration result. Experiments on the CT-MR image registration show that the proposed method can not only greatly improve the registration speed, but also effectively avoid the local convergence and has higher accuracy.

Key words: Fully Convolutional Networks（FCN）, mutual information algorithm, multi-modal, three-dimensional image registration

摘要：

针对传统配准方法在进行三维多模态图像配准时存在收敛速度较慢、容易陷入极值等问题，提出一种基于全卷积神经网络（Fully Convolutional Networks，FCN）和互信息的配准方法。利用FCN模型提取二维图像深层特征并进行粗配准；将得到的配准结果作为互信息算法的初始搜索点，从而使搜索范围缩小至全局最优解附近；利用互信息算法对参数进一步微调优化，得到最优三维配准结果。实验结果表明，在进行CT-MR图像配准时，所提方法不仅可以大幅度提升配准速度，还能有效避免局部收敛的情况，具有更高的准确性。

关键词: 全卷积神经网络, 互信息算法, 多模态, 三维图像配准