Cardiac MRI Reconstruction Using Diffusion Priors and Multi-Scale Transformer-Based Feature Modeling

The advancement of Healthcare 4.0 has driven a growing demand for intelligent cardiac magnetic resonance imaging (MRI) reconstruction to support precision diagnosis and treatment. Currently, while Transformer based reconstruction methods can effectively capture global dependencies in images, they still have some limitations: insufficient high-frequency detail recovery, inconsistent reconstructed structures under undersampling, and poor model interpretability. These issues affect their re liability and practicality in real-world clinical scenarios. This paper proposes a diffusion-based multi-scale feature fusion transformer (DMFT) for cardiac MRI reconstruction, aiming to balance reconstruction accuracy and efficiency. DMFT introduces a compact diffusion latent prior, enhancing the recovery of fine anatomical structures in just eight diffusion iterations. We embed the proposed multi-scale feature fusion (MsFF) module into the Transformer back bone network to further improve feature representation, achieving effective interaction between the latent prior and image features at different spatial scales. This design helps improve the recovery of local details while maintaining global anatomical consistency. The proposed method was evaluated on both the CMR×Recon benchmark dataset and an in-house cardiac MRI dataset. Experimental results at four different acceleration factors show that DMFT consistently achieves superior reconstruction performance com pared to several representative methods. DMFT particularly achieves significant improvements in PSNR, SSIM, and NMSE at 8x and 10x acceleration factors without introducing excessive computational overhead. These results indicate that DMFT provides a promising framework for accurate and efficient cardiac MRIreconstruction, and structure aware prior guidance offers additional interpretability support in Healthcare 4.0 scenarios.