肺部4维CT（4D-CT）图像因受到呼吸、心跳的影响而发生较大的形变，肺内的运动尺度可能大于算法用于优化过程的感兴趣结构（血管、气道等），这可能导致配准算法仅对齐了血管、气道等明显特征。针对肺实质轮廓配准后强度差异性较大的问题，文中提出了以无监督端到端深度学习为基础的多尺度残差可变形肺部CT图像配准算法，使用编码器-解码器结构形式的多尺度深度残差网络作为形变向量场的生成模型，以增强特征表达能力，提高参数利用效率和网络收敛能力；通过多分辨率自注意力融合模块提高网络对多尺度信息的感知能力；设计包含特征校正提取模块的跳跃连接，以有选择地提取编码器输出的特征图，并在重新校准后供解码器学习对齐偏移。最后，在Dir-lab公共数据集上采用文中配准算法与传统算法、目前先进的无监督配准算法进行了比较实验。结果表明：所提出的配准算法在Dir-lab公共数据集上的目标配准误差可以达到1.44 mm ± 1.24 mm，优于传统算法和主流的无监督配准算法；在控制体素折叠率小于0.1%的情况下，估计密集变形向量场耗时小于2.00 s，表明文中算法在对时间敏感的肺部研究中有巨大潜力。

The 4-dimensional CT (4D-CT) images of the lungs undergo large deformations due to respiration and heartbeat, and the scale of motion within the lungs may be larger than the structures of interest (blood vessels, airways, etc.) that the algorithm uses for the optimization process, which may result in the registration algorithms only aligning the obvious features such as blood vessels and airways. To address the problem of high variability of the aligned intensities for structures with large deformations such as the lung parenchyma contour, this paper proposed a multi-scale residual deformable lung CT image alignment algorithm framework based on unsupervised end-to-end deep learning. A multi-scale deep residual network in the form of an encoder-decoder structure was used as a generative model for the deformation field in the proposed registration framework, so as to enhance the feature representation, to increase the effective parameter utilization efficiency parameters and effectively improve the convergence ability of the network. A multi-resolution self-attentive fusion module was used to improve the network’s ability to perceive multi-scale information. And a hopping connection containing a feature correction extraction module was designed to selectively extract the feature maps output by the encoder and recalibrate them for the decoder to learn the alignment offsets. Finally, this paper compared the proposed alignment algorithm with traditional algorithms and the current state-of-the-art unsupervised alignment algorithms on the Dir-lab public dataset. The results show that, the target alignment error of the proposed registration algorithm framework on the Dir-lab public dataset can reach 1.44 mm ± 1.24 mm, which is better than traditional algorithms and the mainstream unsupervised alignment algorithm. In addition, the estimation of the dense deformation vector field takes less than 2.00 s with the control folding voxel less than 0.1%, indicating the great potential of the algorithm in studying time-sensitive lungs.