Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Score-based Generative Priors Guided Model-driven Network for MRI Reconstruction

Xiaoyu Qiao, Weisheng Li, Bin Xiao, Yuping Huang, Lijian Yang

TL;DR

This work tackles MRI reconstruction under undersampling by introducing a three-stage DL workflow that leverages naive score-based Langevin samples as preliminary guidance images (PGIs) without retraining. A denoising module with score-based and cross-domain feature extractors refines these PGIs into denoised guides (DGIs), which in turn steer a densely connected, PGI-guided unrolled network to recover fine details. The approach combines a pretrained score model with a learnable denoising step and a guided reconstruction cascade, achieving state-of-the-art results on knee and FastMRI datasets while requiring fewer training slices and fewer sampling steps. The method offers robustness to distribution shifts and reduces the dependence on extensive diffusion-model tuning, making it practical for clinical MRI pipelines.

Abstract

Score matching with Langevin dynamics (SMLD) method has been successfully applied to accelerated MRI. However, the hyperparameters in the sampling process require subtle tuning, otherwise the results can be severely corrupted by hallucination artifacts, especially with out-of-distribution test data. To address the limitations, we proposed a novel workflow where naive SMLD samples serve as additional priors to guide model-driven network training. First, we adopted a pretrained score network to generate samples as preliminary guidance images (PGI), obviating the need for network retraining, parameter tuning and in-distribution test data. Although PGIs are corrupted by hallucination artifacts, we believe they can provide extra information through effective denoising steps to facilitate reconstruction. Therefore, we designed a denoising module (DM) in the second step to coarsely eliminate artifacts and noises from PGIs. The features are extracted from a score-based information extractor (SIE) and a cross-domain information extractor (CIE), which directly map to the noise patterns. Third, we designed a model-driven network guided by denoised PGIs (DGIs) to further recover fine details. DGIs are densely connected with intermediate reconstructions in each cascade to enrich the information and are periodically updated to provide more accurate guidance. Our experiments on different datasets reveal that despite the low average quality of PGIs, the proposed workflow can effectively extract valuable information to guide the network training, even with severely reduced training data and sampling steps. Our method outperforms other cutting-edge techniques by effectively mitigating hallucination artifacts, yielding robust and high-quality reconstruction results.

Score-based Generative Priors Guided Model-driven Network for MRI Reconstruction

TL;DR

This work tackles MRI reconstruction under undersampling by introducing a three-stage DL workflow that leverages naive score-based Langevin samples as preliminary guidance images (PGIs) without retraining. A denoising module with score-based and cross-domain feature extractors refines these PGIs into denoised guides (DGIs), which in turn steer a densely connected, PGI-guided unrolled network to recover fine details. The approach combines a pretrained score model with a learnable denoising step and a guided reconstruction cascade, achieving state-of-the-art results on knee and FastMRI datasets while requiring fewer training slices and fewer sampling steps. The method offers robustness to distribution shifts and reduces the dependence on extensive diffusion-model tuning, making it practical for clinical MRI pipelines.

Abstract

Score matching with Langevin dynamics (SMLD) method has been successfully applied to accelerated MRI. However, the hyperparameters in the sampling process require subtle tuning, otherwise the results can be severely corrupted by hallucination artifacts, especially with out-of-distribution test data. To address the limitations, we proposed a novel workflow where naive SMLD samples serve as additional priors to guide model-driven network training. First, we adopted a pretrained score network to generate samples as preliminary guidance images (PGI), obviating the need for network retraining, parameter tuning and in-distribution test data. Although PGIs are corrupted by hallucination artifacts, we believe they can provide extra information through effective denoising steps to facilitate reconstruction. Therefore, we designed a denoising module (DM) in the second step to coarsely eliminate artifacts and noises from PGIs. The features are extracted from a score-based information extractor (SIE) and a cross-domain information extractor (CIE), which directly map to the noise patterns. Third, we designed a model-driven network guided by denoised PGIs (DGIs) to further recover fine details. DGIs are densely connected with intermediate reconstructions in each cascade to enrich the information and are periodically updated to provide more accurate guidance. Our experiments on different datasets reveal that despite the low average quality of PGIs, the proposed workflow can effectively extract valuable information to guide the network training, even with severely reduced training data and sampling steps. Our method outperforms other cutting-edge techniques by effectively mitigating hallucination artifacts, yielding robust and high-quality reconstruction results.
Paper Structure (18 sections, 18 equations, 7 figures, 4 tables)

This paper contains 18 sections, 18 equations, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Background
  4. Methods
  5. Score matching and Langevin dynamics
  6. Denoising module
  7. Unrolled network with PGI-guided densely-connected cascades
  8. loss functions
  9. Implementation details
  10. Experiments and Results
  11. Dataset
  12. comparisons with the cutting-edge methods
  13. Ablation Studies
  14. Evaluating the effectiveness of the proposed workflow
  15. Evaluating the network robustness with limited training slices
  16. ...and 3 more sections

Figures (7)

  • Figure 1: Distribution-shift reconstruction results of a pretrained score network on a knee MRI dataset. The score network was trained on a brain MRI dataset. Images in odd columns are ground truth and those in even columns are sampling results.
  • Figure 2: (a) The workflow of the proposed SGM-Net. (b) The network structure of the denoising module and the SIE and CIE blocks. (c) The network structure of the $(k+1)$th GIC.
  • Figure 3: Examples of ground truth images (the first column), undersampled images (the second column) and the corresponding k-space trajectories (the third column) from different datasets.
  • Figure 4: Examples of the reconstructed images of different methods from coronal PD sequence and FastMRI dataset at 4$\times$ and 6$\times$ acceleration. The first rows of each subfigure are the reconstructed images, the second rows are zoomed details in the red square, and the third rows are the error maps corresponding to the ground truth images. The methods from left to right: 1. ground truth; 2. zero-filled; 3. TV; 4. U-Net; 5. D5C5; 6. ISTA-Net; 7. VS-Net; 8. E2EVarNet; 9. ReVarNet; 10. MeDL-Net; 11. vSHARP; 12. SGM-Net (ours).
  • Figure 5: (a) The reconstruction results of the original model at four-fold acceleration. (b) The reconstruction results with 20% sampling steps at four-fold acceleration. The first rows of each subfigure are the reconstructed images, and the second rows are zoomed details in the red squares. The corresponding images from left to right: the ground truth, zero-filled measurements, PGI, DGI, and the final output of $x_T^K$ and $x_z^K$ branch.
  • ...and 2 more figures