LDPM-v2: Towards undersampled MRI reconstruction with multimodal one-step latent diffusion prior.
Authors
Affiliations (5)
Affiliations (5)
- College of Health Science and Environmental Engineering, Shenzhen Technology University, Shenzhen, China. Electronic address: [email protected].
- College of Health Science and Environmental Engineering, Shenzhen Technology University, Shenzhen, China; College of Applied Technology, Shenzhen University, Shenzhen, China.
- School of Electronic Information, Central South University, Changsha, China.
- College of Health Science and Environmental Engineering, Shenzhen Technology University, Shenzhen, China.
- Artificial Intelligence Research Institute, Shenzhen University of Advanced Technology, Shenzhen, China.
Abstract
In recent years, diffusion models have attracted considerable attention in the field of magnetic resonance imaging (MRI) reconstruction, generating high-quality samples through iterative denoising. However, most applications remain constrained by substantial computational overhead in the image domain, error accumulation inherent to iterative inference, and limited controllability over reconstruction details. To address these challenges while leveraging diffusion priors and enhancing adaptation to MR data, we propose a novel method, LDPM-v2: a Latent Diffusion Prior-based framework for undersampled MRI reconstruction guided by multimodal image-and-metadata information. This approach optimizes the text-to-image diffusion priors via a rectified flow strategy and an MRI-tailored variational autoencoder, and further strengthens control over the restoration process using multimodal guidance (including text prompts), enabling high-fidelity, one-step reconstruction. Experiments on the NYU fastMRI brain dataset demonstrate competitive quantitative and qualitative performance at 8× and 10× accelerations. LDPM-v2 achieves robust image-domain and Fourier-domain magnitude consistency and substantially reduced inference time compared with conventional multi-step latent diffusion methods. Additional evaluations demonstrate generalization across the evaluated sampling-pattern, acceleration-factor, and cross-anatomy settings, together with robustness to prompt perturbations and simulated motion corruption.