Image registration is widely used in image-guided lung tumor radiotherapy, but the existing algorithms are not effective against large deformation images. Therefore, this paper proposes an algorithm using a multiscale parallel down sampling module to reduce the image size and obtain a multiscale low-resolution feature map, and the pyramid dilated convolution module is used to extract image features to improve the model’s receptive field. The algorithm adjusts the bias of the neural networks on different deformation features through the adaptive channel attention module to solve the problem in which the model is biased and poor for large and small deformation registration, respectively. Simultaneously, a smoothness constraint is added to the loss function to improve the deformed field smoothness. The training set data-augmentation method is also used to improve the model’s stability and generalization. The proposed algorithm has target registration errors of 1.71 mm and 1.50 mm in the DIR-lab and Creatis datasets, respectively, whereas the one-iteration fully convolutional neural network (FCN) has target registration errors (TREs) of 2.83 mm and 2.01 mm in the above datasets, respectively. The experimental results show that the TRE of the algorithm is significantly smaller than that of the FCN algorithm, and the generalization and stability performances of the algorithm are also improved.