【摘要】 三维肺部电子计算机断层扫描(computed tomography,CT)图像非刚性配准是医学图像配准领域中最重要的任务之一.但是,肺部组织受呼吸运动影响而产生的非线性形变与大尺度位移给三维肺部CT图像的非刚性配准带来巨大挑战.针对这一难题,设计开发了一种基于无监督学习端到端的配准方法.通过改进现有U-Net神经网络结构,在跳接之间引入Inception模块,充分融合多尺度深层特征生成高精度的稠密位移向量场.为保证位移向量场光滑,在损失函数中加入雅可比正则化项,以达到训练中显式惩罚位移向量场中奇点的目的.另外,为缓解现有公开数据资源有限导致的过拟合问题,提出了一种基于三维薄板样条(3D-thin plate spline,3D-TPS)变换的数据增强方法实现对训练数据的扩充,将具有60套三维肺部CT图像的训练数据集EMPIRE10扩充为6 060套以满足卷积神经网络训练的需要.设计验证实验,通过与基于学习的Voxelmorph方法和两个包含传统方法配准工具包ANTs和Elastix进行比较.实验结果表明:在公开可用的DIR-Lab 4 DCT数据集上,所提出的方法在目标配准误差(target registration error,TRE)上达到次优的2.09 mm,平均Dice得分达到最优的0.987,同时所生成的扭曲图像中几乎不存在折叠体素.更多还原

【Abstract】 Deformable registration of 3D lung CT images is crucial in medical image registration. However,nonlinear deformation and large-scale displacement of lung tissues caused by respiratory motion pose great challenges in the deformable registration of 3D lung CT images. Thus,we present a fast end-to-end registration method based on unsupervised learning. We optimized the classic U-Net model and added Inception modules between skip connections.The Inception module aims to capture and merge information at different spatial scales for generating a high-precision dense displacement vector field. To ensure a smooth displacement vector field,we introduced the Jacobian regularization term into the loss function to directly penalize the singularity of the displacement field during training. The existing publicly available datasets cannot implement model training. To address over-fitting caused by limited data resources and to expand the training data,we proposed a data augmentation method based on a 3D thin plate spline transform. Moreover,6 060 CT scans will be generated based on the EMPIRE10 dataset,which contains 60 original CT scans to meet the requirement of convolution neural network training. Regarding the DIR-Lab 4 DCT dataset,we achieved a target registration error of 2.09 mm,an optimal Dice score of 0.987,and almost no folding voxels in comparison with the experimental results obtained using the deep learning method Voxelmorph and registration packages,such as advanced normalization tools(ANTs) and Elastix.更多还原