Research on Medical Image Super-Resolution Reconstruction Using a Transformer‑RRDB Hybrid Generator

DOI: https://doi.org/10.62517/jike.202604316

Author(s)

Yingjie Song

Affiliation(s)

School of Electrical and Information Engineering, Heilongjiang Institute of Technology, Jixi, Heilongjiang, China

Abstract

The limited resolution of magnetic resonance imaging (MRI) is one of the important factors affecting the accuracy of clinical diagnosis. This paper proposes a medical image super-resolution reconstruction method based on a Transformer-RRDB hybrid generator, targeting the 4×super-resolution task of brain T2-weighted MRI images. The method employs a cascade structure of residual dense blocks and Swin Transformer, where RRDB is used to extract local texture features and the window-based self-attention mechanism models the global spatial correlations of brain anatomical structures. In the design of the loss function, we combine the pixel-wise L1 loss, perceptual loss, frequency-domain amplitude-phase joint constraint, and DDO implicit discriminative regularization with a fixed scoring proxy, forming multi-dimensional supervisory signals. The training adopts a two-stage transfer learning strategy of pre-training on DF2K natural images followed by fine-tuning on IXIT2 medical images. Experimental results show that the proposed method achieves PSNR 28.42 dB, SSIM 0.845, LPIPS 0.107, and gEn 0.092 on the IXIT2 test set, outperforming comparative methods such as Bicubic, SRCNN, ESRGAN, and SwinIR in all metrics, which validates the effectiveness of the hybrid architecture and frequency-domain constraints in medical image super-resolution.

Keywords

Medical Image Super-Resolution; Transformer; RRDB; Frequency-Domain Loss; Transfer Learning

References

[1]Huang W F, Liao X Y, Chen H, et al. Deep local-to-global feature learning for medical image super-resolution. Computerized Medical Imaging and Graphics, 2024, 115:102374. [2]Li G Y, Lv J, et al. Multicontrast MRI Super-Resolution via Transformer-Empowered Multiscale Contextual Matching and Aggregation. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(9):12004-12014. [3]Liu Q H, Zhang W K, Zhang Y T, et al. DGEDDGAN: A dual-domain generator and edge-enhanced dual discriminator generative adversarial network for MRI reconstruction. Magnetic Resonance Imaging, 2025, 119:110381. [4]Dong C, Loy C C, He K M, Tang X O. Image super-resolution using deep convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016, 38(2): 295-307. [5]Ledig C, Theis L, Huszár F, et al. Photo-realistic single image super-resolution using a generative adversarial network//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017: 4681-4690. [6]Wang X, Yu K, Wu S, et al. ESRGAN: Enhanced super-resolution generative adversarial networks//Proceedings of the European Conference on Computer Vision (ECCV) Workshops. Springer, 2018: 63-79. [7]Nandal P, Pahal S, Khanna A, et al. Super-resolution of medical images using Real ESRGAN. IEEE Access, 2024, 12: 176155-176170. [8]Sun J W, Li Z Y, Li P C, et al. Improving the diagnostic performance of computed tomography angiography for intracranial large arterial stenosis by a novel super-resolution algorithm based on multi-scale residual denoising generative adversarial network. Clinical Imaging, 2023, 96: 1-8. [9]Liang J Y, Cao J Z, Sun G L, et al. SwinIR: Image Restoration Using Swin Transformer//Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, 2021: 1833-1844. [10]Li J L, Wang G Z, Zhao H W. Magnetic resonance image super-resolution reconstruction based on frequency‑domain constraints and cross-fusion feature. Chinese Journal of Medical Physics, 2023, 40(1): 31-38.