SwinMED: Edge-Texture Enhanced SwinIR for Medical Image Super-Resolution

Abstract

Medical image super-resolution (SR) addresses a fundamental clinical challenge: hardware constraints, acquisition trade-offs, and patient motion frequently produce MRI images with insufficient spatial detail, compromising lesion boundary delineation and the detectability of small pathological findings. Although Transformer-based architectures-and SwinIR in particular-have demonstrated strong SR performance on natural photographic datasets, they do not incorporate explicit mechanisms for enhancing the edge and texture features most relevant to clinical MRI interpretation, creating a domain gap that limits their direct applicability to medical imaging. In this study, we propose SwinMED, a lightweight extension of SwinIR that integrates a learnable Edge-Texture (ET) module (similar to 110K parameters) targeting multi-scale edge recovery and local texture enhancement through parallel Sobel branches, dilated texture convolutions, gated fusion, and channel attention. The ET module is optimized via a self-supervised back-projection protocol that requires no paired high-resolution reference images, making SwinMED directly applicable in clinical settings. The framework was evaluated on four prostate MRI datasets (KEAH, ProstateX, I2CVB, Prostate158) across T2-weighted, dynamic contrast-enhanced (DCE), and diffusion-weighted (DWI) sequences, under two degradation scenarios and three upscaling factors (& times;2, & times;3, & times;4), covering 72 independent experimental conditions. Quantitative evaluation using PSNR, SSIM, MS-SSIM, and LPIPS shows consistent improvements over the baseline SwinIR configuration across all conditions, with average gains of +4.03 dB in PSNR, +0.114 in SSIM, +0.041 in MS-SSIM, and a-0.201 reduction in LPIPS. Positive PSNR and SSIM gains were observed in all 72 conditions; LPIPS improved in 70 out of 72 conditions. These findings indicate that lightweight, domain-targeted edge-texture refinement can complement pretrained transformer SR backbones in medical imaging contexts, providing consistent reconstruction improvements without retraining the full model and without requiring paired high-resolution training data.