MSwinFuSR: a multibranch transformer network for MRI image super-resolution

To address the challenges posed by the complex distribution and difficulty in detail restoration of medical images, this paper proposes a deep network model with multiple structural enhancements based on the SwinFuSR super-resolution network, aiming to improve the performance of super-resolution reconstruction for magnetic resonance imaging (MRI). The proposed model incorporates three key innovations. First, a feature mapping module is designed by integrating Laplacian pyramid decomposition with multi-scale Kan convolution, which effectively reduces the complexity of raw MRI images and enhances network training efficiency. Second, a Biformer structure is introduced to operate in parallel with the ACDF module for feature extraction, thereby strengthening the modeling capacity for regional structural information. Finally, a fusion module is constructed to reinforce the local perception and representation capabilities of features extracted by both ACDF and Biformer modules. Experiments conducted on the FastMRI brain medical image dataset demonstrate that the improved model, MSwinFuSR, achieves a PSNR of 34.43 dB and an SSIM of 0.9205, representing improvements of 1.54 dB and 0.0389, respectively, over the original network (32.89 dB, 0.8816), and significantly outperforming other mainstream super-resolution methods.