<p>Anisotropic magnetic resonance imaging (MRI) has limitations hindering accurate anatomical characterization and clinical diagnosis. Therefore, it is crucial to reconstruct anisotropic MRI super-resolution (SR) into isotropic high-resolution MRI. Multi-modal MRI SR reconstruction methods, which restore high-quality MRI images from degraded target modality under the guidance of reference modality, perform better than monomodal SR methods. However, existing convolutional neural networks (CNNs) based multi-modal MRI SR methods lack simple and effective methods to fuse information from different modalities fully. Although transformer has achieved remarkable results in computer vision tasks, the research on multi-modal 3D MRI SR reconstruction is relatively scarce and still has room for improvement such as low accuracy of the model, high memory requirements, and poor generalization capacity. For these reasons, we propose a simple and efficient multi-modal feature transfer network (MFTNet) for anisotropic 3D MRI image SR reconstruction. Specifically, this network mainly consists of a feature transfer sub-network (FTSN) and a reconstruction branch for the target modality. The former is used to transfer useful high-frequency information from the reference modality, and the latter is used to fully fuse the transferred features and reconstruct high-quality images. Sufficient experiments on three publicly available datasets have shown that MFTNet is superior to state-of-the-art (SOTA) monomodal and multi-modal SR methods, and exhibits strong generalization ability. At the same time, a good balance has been achieved between model performance and cost. Further, the evaluation results on real clinical data show that MFTNet has better performance than SOTA, demonstrating its great potential for clinical applications.</p>