<p>Medical image registration plays a critical role in tumor growth monitoring, radiation therapy and disease diagnosis. Although current medical image registration methods can achieve considerable results, they are easily influenced by noisy regions during the processing of image data with large deformations, making it difficult to maintain the continuity of the deformation field and achieve the desired registration accuracy. To address these issues, this research proposes a dual-stream network combining multiple attention mechanisms for deformable medical image registration (MADS-Net). The model utilizes a dual-encoder structure to enhance registration performance through the cross-perceptual attention (CPA) module and the multi-dimensional attention (MDA) module. Specifically, unlike traditional methods that concatenate fixed and moving images as input to the encoder, MADS-Net employs dual encoders to separately extract features, which facilitates explicit feature matching between image pairs. The efficient feature extraction module, CPA, embedded in the encoder, is beneficial for extracting long-range global features, while the MDA module embedded in the skip connections filters redundant features and extracts local refined features to improve the U-Net. To validate the performance of our proposed network, we compared it with other methods on the LPBA40, IXI and OASIS public datasets. The experimental results demonstrate that MADS-Net exhibits superior registration performance under commonly used evaluation metrics.</p>