CobDiffMRI: a cross-conditional diffusion model for removing artefacts in undersampled magnetic resonance imaging

As a medical imaging method, magnetic resonance imaging (MRI) can provide good soft tissue contrast, but it is necessary to accelerate MRI scanning due to the long reconstruction time which results in serious artefacts of undersampled k-space data reconstruction. Existing methods still have various limitations in image fidelity and dependence on full sampled acquisition. This article proposes a reconstruction method based on cross-conditional diffusion model CobDiffMRI to address the issue of undersampled reconstruction artefacts, using undersampled k-space signals and artefact images as conditions, applying the conditional diffusion model to the training and reconstruction process of MRI imaging in which learn the diffusion process and image features through training. At the same time, a cross-regularisation loss function is designed to predict noise more reasonably and effectively. With a diffusion step of 2000, the artefacts caused by insufficient features under undersampled conditions is well optimised, which makes the goal clear and the features distinguishable, providing a new idea for the application of neural networks in MRI.