Highly Undersampled MRI Reconstruction via a Single Posterior Sampling of Diffusion Models

Jin Liu, Qing Lin, Zhuang Xiong, Shanshan Shan, Chunyi Liu, Min Li, Feng Liu, G. Bruce Pike, Hongfu Sun, Yang GaoAuthorea 2025 · DOI

TL;DR

This paper introduces SSDM-MRI, a novel diffusion model-based framework for Magnetic Resonance Imaging (MRI) reconstruction that achieves high-quality results from highly undersampled data in a single posterior sampling step. By employing an iterative selective distillation technique and a shortcut reverse sampling strategy, SSDM-MRI significantly reduces inference time, making it comparable to traditional deep learning methods while maintaining superior performance.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a crucial, non-invasive diagnostic tool, but its lengthy acquisition times can lead to patient discomfort and motion artifacts. Accelerating MRI scans by undersampling k-space (the raw data collected by the scanner) is a key research area. While traditional methods and early deep learning approaches struggle with high acceleration factors (e.g., 8x or more), recent diffusion models (DMs) show promise. However, a major bottleneck for DMs is their extremely slow inference due to a large number of iterative sampling steps required to generate an image. This paper aims to bridge this gap by developing a DM-based reconstruction method that is both highly accurate and computationally efficient, specifically addressing the long inference time issue.

Methods

The core innovation of SSDM-MRI lies in its approach to accelerating the diffusion model's reverse sampling process. Instead of the typical hundreds or thousands of steps, SSDM-MRI achieves reconstruction in a single posterior sampling step. This is accomplished through a two-stage process:

  1. Conditional Diffusion Model Pretraining: A standard conditional diffusion model is trained in a supervised manner. This model learns to generate clean MRI images from noisy inputs, guided by the undersampled k-space data and a zero-filling reconstruction. This stage requires a significant number of reverse steps (T=128 in this work).
  2. Iterative Selective Distillation with Shortcut Reverse Sampling: This is the key to acceleration. The pretrained model is progressively "distilled" into a faster student model over multiple iterations (K=4 in this work).
  3. Iterative Selective Distillation: Instead of distilling the entire reverse process, this technique focuses only on the latter half (T/2) of the sampling steps in each distillation iteration. This selective approach is more efficient and less prone to performance degradation.
  4. Shortcut Reverse Sampling: To further accelerate the process, the reverse sampling doesn't start from pure noise but from a "latent point" derived from the zero-filling reconstruction. This shortcut path significantly reduces the number of steps needed.
  5. Data Consistency (DC): A projection-based data consistency step is applied at the end to ensure the reconstructed image adheres to the acquired k-space measurements.

This combination effectively "compresses" the diffusion model's generative capability into a single-step inference process.

Datasets

The study utilizes both publicly available and in-house datasets for training and evaluation:

  • fastMRI Brain Dataset: A large dataset of 2D brain images (320x320) used for training and evaluation. Various undersampling masks (1D Gaussian, 2D Poisson) were simulated at different acceleration factors (AFs) from 4x to 15x.
  • fastMRI Knee Dataset: Used for out-of-distribution (OOD) evaluation to test generalization capabilities.
  • In-house mGRE Brain Dataset: A multi-echo gradient echo dataset used for training and evaluating Quantitative Susceptibility Mapping (QSM) acceleration. This dataset consists of complex-valued slices.
  • Multi-coil Datasets: Both retrospective and prospective multi-coil brain and knee datasets were used for evaluation.

Key Insights and Ideas

  • Single-Step Diffusion for MRI: The most significant insight is demonstrating that diffusion models, known for their generative power but also their slowness, can be adapted for high-quality MRI reconstruction in a single inference step. This overcomes a major practical hurdle for DM adoption in clinical settings.
  • Selective Distillation is Key: Focusing distillation on only the latter half of the diffusion process (T/2) is more effective and stable than distilling the entire process. This suggests that the later stages of diffusion are more critical for fine-tuning the image details.
  • Shortcut Sampling for Efficiency: Starting the reverse process from a pre-computed latent point (rather than pure noise) significantly reduces the required sampling steps, further accelerating inference.
  • Trade-off Between Performance and Speed: The paper acknowledges that while more distillation steps generally lead to better performance, there's a point of diminishing returns. The chosen four distillation steps strike a good balance between reconstruction quality and computational efficiency.
  • Generalization Beyond Training Data: SSDM-MRI shows promising generalization capabilities, performing well on OOD knee data even when trained primarily on brain images, suggesting it learns underlying anatomical priors.

Key Results

  • Superior Reconstruction Quality: SSDM-MRI consistently outperformed other state-of-the-art methods (including traditional iterative methods like P-LORAKS and CSPC, and other deep learning/DM methods like MoDL, U-net, Score-MRI, and CCDF) in terms of numerical metrics (PSNR, SSIM), error maps, and preservation of fine image details.
  • Exceptional Speed: The reconstruction time for a 320x320 brain slice is approximately 0.45 seconds, which is comparable to simple U-net architectures and orders of magnitude faster than other DM-based methods (which can take minutes).
  • Accurate QSM Reconstruction: The method demonstrated excellent performance in Quantitative Susceptibility Mapping (QSM) acceleration, preserving latent susceptibility information and achieving accurate quantitative measurements in deep grey matter regions.
  • Robustness to Acceleration: SSDM-MRI showed good performance across a range of acceleration factors (AF=4x to 15x) with minimal artifacts.
  • Stable Reconstructions: Uncertainty analysis indicated that SSDM-MRI provides more stable and robust reconstructions compared to Score-MRI and CCDF, especially at higher acceleration factors.

Limitations and Future Work

  • Number of Distillations: While four distillation steps were found to be optimal, the paper notes that performance slightly decreases with more distillations, indicating a trade-off between model performance and computational cost. Future work could explore optimizing this trade-off further.
  • Data Consistency Methods: The study used a simple projection-based data consistency (PDC) method. While effective, the authors suggest that more advanced DC methods (like two-step DC or learning-based DC) could potentially further improve results, though the gains might be marginal compared to the initial DC incorporation.
  • Generalization to Different Modalities/Applications: While SSDM-MRI showed good generalization to OOD knee data, further exploration into its applicability for other MRI contrasts or 3D reconstruction could be beneficial.
  • Artifacts at Very High AFs: Although generally robust, the paper acknowledges that fine details can become blurred at extremely high acceleration factors, suggesting ongoing work to improve performance in such scenarios.

Summary updated

Abstract

Incoherent k-space undersampling and deep learning-based reconstruction methods have shown great success in accelerating MRI. However, the performance of most previous methods will degrade dramatically under high acceleration factors, e.g., 8× or higher. Recently, denoising diffusion models (DM) have demonstrated promising results in solving this issue; however, one major drawback of the DM methods is the long inference time due to a dramatic number of iterative reverse posterior sampling steps. In this work, a Single Step Diffusion Model-based reconstruction framework, namely SSDM-MRI, is proposed for restoring MRI images from highly undersampled kspace. The proposed method achieves one-step reconstruction by first training a conditional DM and then iteratively distilling this model four times using an iterative selective distillation algorithm, which works synergistically with a shortcut reverse sampling strategy for model inference. Comprehensive experiments were carried out on both publicly available fastMRI brain and knee images, as well as an in-house multi-echo GRE (QSM) subject. Overall, the results showed that SSDM-MRI outperformed other methods in terms of numerical metrics (e.g., PSNR and SSIM), error maps, image fine details, and latent susceptibility information hidden in MRI phase images. In addition, the reconstruction time for a 320×320 brain slice of SSDM-MRI is only 0.45 second, which is only comparable to that of a simple U-net, making it a highly effective solution for MRI reconstruction tasks.
BibTeX
@misc{liu2025highly,
  author = {Jin Liu and Qing Lin and Zhuang Xiong and Shanshan Shan and Chunyi Liu and Min Li and Feng Liu and G. Bruce Pike and Hongfu Sun and Yang Gao},
  title = {{Highly Undersampled MRI Reconstruction via a Single Posterior Sampling of Diffusion Models}},
  year = {2025},
  doi = {10.22541/au.176253018.86072169/v1},
  howpublished = {Authorea},
}
Figure