Autocalibration signal-guided multi-frequency diffusion model for dynamic MRI reconstruction

Abstract

Objective. The autocalibration-signal (ACS) region plays a pivotal role in dynamic magnetic resonance imaging (MRI) reconstruction, yet its full potential remains under-exploited in current generative frameworks, which rely mainly on data-driven priors while overlooking ACS-embedded physical information. This work aims to propose a novel multi-frequency aware diffusion model guided by structural priors derived from the ACS region to enhance the quality of dynamic MRI reconstruction. Approach. Specifically, the intrinsic phase consistency and spatial smoothness within the ACS region are leveraged to construct a robust prior that stabilizes the diffusion trajectory. This prior further guides the restoration of high-frequency components, thereby improving temporal coherence and structural fidelity. Furthermore, the non-ACS region is divided into two complementary frequency bands, each jointly modeled with the ACS region to establish frequency-conditioned diffusion trajectories across the frequency domain. To further refine reconstruction performance, we incorporate low-rank regularization across time frames and enforce data-consistency constraints, which effectively suppress motion artifacts while preserving fine anatomical details. Main results. Quantitative evaluations demonstrate that the proposed method achieves competitive reconstruction quality across various sampling patterns and acceleration factors. Under Radial sampling at an acceleration factor of R= 8, it attains PSNR of 32.53 ± 0.34 dB, NMSE of 0.011 ± 0.001 and SSIM of 0.8583 ± 0.0057, outperforming all competing methods. Notably, the proposed method achieves competitive performance comparable to supervised models without requiring any fully-sampled data or external training labels. Significance. The proposed method enables high-quality dynamic MRI reconstruction by exploiting intrinsic ACS information, offering a practical solution without fully-sampled training data.