Deformable medical image registration research, crucial in image analysis and clinical fields, often relies on ConvNets to create U-shaped networks. However, due to the limited receptive fields of ConvNets, traditional methods may be less effective in aligning images with substantial dissimilarity. This paper presents a hybrid approach merging ConvNet with attention mechanism, leveraging the remote modelling abilities of attention mechanism to address the above challenges. First, we design a multiscale transformer fusion module (MTF) and integrate it into the skip connections of the U-shaped network, enhancing spatial relationship modelling. Second, We propose a plug-and-play registration head for positional reference in deformation field generation. Additionally, to enhance the lung registration accuracy in chest radiographs, we design a new loss function named lung region loss. Compared to current methods, this hybrid approach shows marked improvements in average Dice Similarity Score (DSC), Hausdorff Distance (HD), and Average Symmetric Surface Distance (ASSD) across three benchmark datasets, demonstrating its effectiveness in medical image registration.