荧光分子断层成像与计算机断层成像的2D/3D配准方法研究

2D/3D registration method between fluorescence molecular tomography and computed tomography

针对荧光分子断层成像(fluorescence molecular tomography,FMT)中的二维光学图像与三维计算机断层成像(computed tomography,CT)图像在维度和风格等方面的差异导致配准困难的问题,本研究提出了一种基于任务回归粗配准及梯度下降精配准的两阶段配准方法。在粗配准阶段,本研究提出了目标姿态预测模型RR-Net,实现对初始位姿参数的快速估计;在精配准阶段,基于可微成像渲染模块,在粗配准参数的基础上进行迭代训练,进一步提高配准精度。本研究在公开小鼠数据集及实验室自采数据共238140张轮廓图像中验证了该配准方法的有效性。实验结果显示,本研究提出的2D/3D配准方法在FMT光学图像和CT图像配准中,相似系数为0.96±0.03,平均目标配准误差为(1.14±0.83)mm。该方法的配准效果及稳定性均优于传统方法。

Considering the difficult registration due to the differences in dimension and style between fluorescence molecular tomography(FMT)optical images and computed tomography(CT)images,we proposed a two-stage registration method based on task regression coarse registration and gradient descent fine registration.In the coarse registration stage,a target pose prediction model RR-Net was utilized to quickly estimate the initial pose parameters.In the fine registration stage,a differentiable imaging rendering module was employed to iteratively train based on the coarse registration results,rapidly converging to precise registration results.The experiments were conducted on 238140 contour images from public data and laboratory-collected data to validate the effectiveness of the proposed registration method.The experimental results showed that the proposed 2D/3D registration method in the registration of FMT optical images and CT images registration,the similarity coefficient and average target registration error could reached 0.96±0.03,(1.14±0.83)mm,respectively.The registration effect and stability of this research are better than that of the traditional methods.