基于神经网络的双X射线影像2D-3D配准算法-《中国医学物理学杂志》

文章信息/Info

Title:: 2D-3D double X-ray image registration method based on neural network

Author(s):: SHEN Yanyan1; FENG Hansheng2; 1. School of Electrical Engineering and Automation, Anhui University, Hefei 230601, China; 2. Institute of Plasma Physics, Chinese Academy of Sciences, Hefei 230031, China

Keywords:: Keywords: double X-ray image; 2D-3D registration; convolutional neural network; geometric decomposition

摘要:: 针对基于迭代优化的传统2D-3D医学图像配准算法运行速度慢，难以达到实时配准的要求，本研究提出一种实时2D-3D配准方法。通过将空间刚体变换参数分解到两个平面上，将2D-3D配准简化为两个步骤，包含2D-2D近似刚体配准与单参数2D-3D刚体配准。同时利用深度卷积神经网络拟合患者X射线影像残差与其对应姿态差异间的非线性映射关系，从X-DRR图像对的残差回归出空间刚体变换参数。经由头颅CT数据训练后的网络，在0.04 s内完成了高精度的双X射线配准。本研究提出的配准方法满足了放疗过程中进行实时2D-3D配准工作的要求。

Abstract:: Abstract: Based on the situation that the traditional 2D-3D medical image registration algorithm based on iterative optimization can not realize real-time registration due to slow running speed, a real-time 2D-3D registration method is proposed. By decomposing spatial rigid transformation parameters into two planes, 2D-3D registration is simplified into two steps, including 2D-2D approximate rigid registration and single-parameter 2D-3D grid registration. Meanwhile, deep convolutional neural network is used to fit the nonlinear mapping between X-ray images residual and its corresponding postural difference, and the space rigid transformation parameters are regressed from the residual of the X-DRR image pair. The network trained by head CT data can complete the high-precision double X-ray registration within 0.04 s. The proposed registration method can satisfy the requirements of real-time 2D-3D registration during radiotherapy.