Wavelet Dual-Stream Network for Brain MR Image Super-Resolution.

High-resolution (HR) magnetic resonance (MR) images provide more detailed information for reliable diagnoses and quantitative medical image analyses. Deep convolutional neural networks (CNNs) have demonstrated their ability to effectively retrieve HR MR images from low-resolution (LR) MR images. However, most CNN-based super-resolution (SR) algorithms treat content and background information equally, ignoring the unique properties of MR images, such as low contrast, intricate tissue textures, and sparse backgrounds. We present a Wavelet Dual-Stream Network (WDN) for accurate MR SR that addresses the issues raised above. First, a wavelet transform is leveraged at the network's beginning to decouple the inputs, which are divided into high-frequency and low-frequency sub-bands. The high-frequency sub-bands relate to content information with larger frequency changes, and the low-frequency sub-bands correspond to background information. In addition, we devise a two-branch structure to reconstruct the high-frequency and low-frequency features separately. On the one hand, we design the U-Net Attention (U-A) mechanism for focusing the network's attention on regions with critical information. On the other hand, due to the correlation between high-frequency and low-frequency branches, We establish the Cross Attention Block (CAB) to accomplish the interaction between two branches. CAB takes advantage of the redundancy of information between different branches to distill the information of the current branch. Finally, the inverse wavelet transform is utilized to couple the modified high-frequency and low-frequency sub-bands as SR. Extensive experiments confirm the effectiveness of the WDN, which provides a clear improvement over the state-of-the-art method in both subjective and objective evaluations.