Aliased Reference Diffusion Reconstruction Strategy for Accelerated MR Imaging and Fine-tuning Performance Studies

Diffusion models guided by stochastic differential equations (SDEs) have exhibited superior performance in image recovery. However, current diffusion-based methods typically necessitate large iterations to recover images, which significantly imposes the computational burden. To solve this problem, we propose a novel reconstruction method called Aliased Reference Diffusion Reconstruction Strategy (ARDRS). ARDRS is trained by mean-shift SDE to complete priori learning of aliasing process, and accomplishes reconstruction of the ground truth starting from the aliased image. Secondly, we investigate the reconstruction performance of diverse normalization methods for determining the optimal range. Finally, further analyses about fin-tuning sampling steps and noise scale are conducted to ascertain the effects toward reconstruction. Experiments are conducted on publicly available fastMRI dataset. Results show that main texture of MR images can be distinct reconstructed at the acceleration rate of 8. Meantime, we provide a fine-tuning approach without changing model, which can achieve same reconstructions when the data pre-processing is changed.