Image reconstruction based on a physics-informed reverse diffusion model trained with magnitude-only data

We trained a score-based generative diffusion model with cardiac MR images, which allows generating new, randomized instances of the given data distribution. By conditioning each step of the underlying reverse time stochastic differential equation with a physics-informed data consistency step, undersampled MR data can be reconstructed. An initial estimation of the complex phase, which slowly transfers into the actual phase of the image, allows to train the diffusion model with magnitude data only. The approach was evaluated in fast spiral dynamic cardiac MRI at 1.5T, where it provided superior SNR-levels compared to alternative acceleration techniques.