A Swin Transformer-Based Cold Diffusion Model for Accelerated MRI Reconstruction

Clinical diagnosis has become easier because of magnetic resonance imaging (MRI) since detailed images of soft tissues are possible. While having its own advantages MRI has multiple limitations. MRI machines are generally expensive, and extended scan time can (over)burden clinical healthcare systems at the expense of medical work efficiency. The rise of deep learning technologies has spurred the development of methods to accelerate MRI reconstruction. In this regard, we introduce a novel cold-diffusion-based approach to accelerate MRI reconstruction. This method leverages the Swin Transformer architecture for effective feature extraction and incorporates a high-frequency self-attention mechanism to improve the recovery of high-frequency components, thus enhancing image quality. Unlike traditional diffusion models that rely on Gaussian noise, our cold-diffusion model simulates data loss through specific degradation processes like Cartesian undersampling. Experimental results demonstrate that the Swin-CDMR approach outperforms existing techniques in reconstruction quality, achieving superior structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) scores while maintaining robustness across varying acceleration factors (AF).