Abstract:

Acquiring high-resolution (HR) magnetic resonance (MR) images remains challenging in clinical practice because of limitations in scanning conditions. As an effective post-processing technique for enhancing image resolution, super-resolution (SR) plays an important role in improving the spatial resolution of MR images and supporting clinical diagnosis. Most existing MR image SR methods use deep neural networks to learn a direct mapping from low-resolution (LR) images to HR images. However, LR images lose high-frequency information during the degradation process, making it difficult to reconstruct fine textures and details using only the limited information available in LR images. To improve the accuracy of detail and texture reconstruction, this study introduces codebook priors into MR image SR. The codebook is trained to learn prior information from HR images and is used to guide the reconstruction of LR images. Specifically, we propose a two-stage High-Quality Codebook Prior Network (CPNet). In the first stage, an encoder-decoder network is pretrained on HR images to extract high-quality codebook priors. In the second stage, a Cross-Channel Attention Fusion Module (CCAFM) is developed to incorporate the codebook priors into the SR network. In addition, a novel Gated Convolutional Transformer Block is designed as the basic building block to enhance feature extraction. Experimental results demonstrate the effectiveness of the proposed method and show that it achieves state-of-the-art performance. In particular, the introduction of codebook priors substantially improves the network's ability to restore fine textures and details.

Key words: super-resolution, magnetic resonance image, codebook priors