NRD-Net: Non-local Residual Dense Network for Brain MR Image Super-Resolution

Super-resolution can significantly enhance image visibility and restore image features without requiring scanning devices to be updated. It is notably useful for magnetic resonance imaging (MRI), which suffers from low-resolution issue. In practice, MR images possess more intricate texture details than natural images, leading to the issue that existing super-resolution algorithms struggle to reach acceptable performance, particularly for brain MR images. To this end, we propose a non-local residual dense network (NRD-Net) for brain MR image super-resolution. In NRD-Net, shallow features are first extracted using a convolutional layer. Next, we propose to adaptively weight the extracted features using a non-local residual dense block, which captures the long-range relationship between features and enables the network to incorporate global information while retaining rich deep features. Finally, HR images are reconstructed using the reconstruction block based on atrous spatial pyramid pooling and sub-pixel convolution. Extensive experiments illustrate that, compared with existing super-resolution approaches, our NRD-Net provides better reconstruction performance with promising peak signal-to-noise ratio and structural similarity, as well as better anatomical structural details.