PDDR: Diffusion model speeds up cardiac MRI reconstruction 2x

Real-time cardiac cine MRI captures the beating heart during free breathing, but severe undersampling and motion make high-quality reconstruction extremely challenging. Existing methods either use simple priors that lose detail or heavy diffusion models that are computationally prohibitive. Researchers Florian Fürnrohr and Reinhard Heckel from the Technical University of Munich introduce Piecewise Dynamic Diffusion Regularization (PDDR), a reconstruction method that integrates a spatiotemporal diffusion model as a generative prior within a variational framework. The model uses dedicated spatial layers for anatomical structure and temporal layers for cardiac motion, learned from gated cine data. PDDR applies the dynamic prior in a piecewise manner, allowing the efficient use of spatiotemporal diffusion models on long real-time sequences without excessive memory or compute.

Experiments on both retrospectively accelerated and prospective real-time cine MRI demonstrate that PDDR outperforms classical, unsupervised, and other diffusion-based methods, delivering high-quality reconstructions with substantially reduced computation time. This makes it a practical and scalable solution for free-breathing, real-time cardiac MRI in clinical settings. The code is publicly available on GitHub, enabling others to replicate and build upon the work. The method's ability to balance image quality and speed addresses a key bottleneck in deploying advanced MRI reconstruction techniques in routine practice.