OS-DDPM: One-Step Denoising Diffusion Probabilistic Model for Anisotropic MRI Super-Resolution

In magnetic resonance imaging (MRI), anisotropic volumes with low through-plane resolution are typically acquired. Recently, diffusion models have shown strong performance in anisotropic MRI super-resolution (SR). However, iterative sam-pling limits the clinical applicability of diffusion models. To address this problem, we propose a One-Step Denoising Diffusion Probabilistic Model (OS-DDPM) for anisotropic MRI SR, which can generate isotropic High-Resolution (HR) data in a single sampling step. Firstly, we construct a Dual-Domain One-Step Generator (DDOS-Generator) comprising a student network for initial reconstruction in the spatial domain and a wavelet denoising module for noise suppression and detail refinement in the wavelet domain. Secondly, variational score distillation is applied to distill the generative prior from the pre-trained multi-step DDPM (teacher) to OS-DDPM, which can produce high-quality one-step SR data comparable to multi-step SR data. Finally, diffusion-based noise-aware contrastive learning is designed to bridge the distribution mismatch between one-step SR data and HR data. Extensive experiments on two public datasets demonstrate that the proposed method offers a fast inference speed and outperforms existing anisotropic SR methods and its teacher diffusion model in most metrics.