Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Priors for MRI Reconstruction Trained from Magnitude-Only Images Using Phase Augmentation

Guanxiong Luo, Xiaoqing Wang, Mortiz Blumenthal, Martin Schilling, Erik Hans Ulrich Rauf, Raviteja Kotikalapudi, Niels Focke, Martin Uecker

TL;DR

This work addresses accelerating MRI reconstruction by learning priors from magnitude-only images. It introduces phase augmentation via a diffusion-based prior to synthesize realistic complex-valued images, enabling six priors trained on varying dataset sizes and complex/magnitude data to regularize linear and nonlinear reconstructions. The study demonstrates that priors trained on complex-valued data outperform magnitude-only priors, with larger datasets offering greater robustness; diffusion priors further improve 3D reconstructions and often exceed conventional $\ell_1$-wavelet regularization under high undersampling. Practically, phase augmentation provides a scalable workflow to leverage existing magnitude-only image databases for robust, phase-informed MRI reconstruction across diverse undersampling schemes and coil configurations.

Abstract

Purpose: In this work, we present a workflow to construct generic and robust generative image priors from magnitude-only images. The priors can then be used for regularization in reconstruction to improve image quality. Methods: The workflow begins with the preparation of training datasets from magnitude-only MR images. This dataset is then augmented with phase information and used to train generative priors of complex images. Finally, trained priors are evaluated using both linear and nonlinear reconstruction for compressed sensing parallel imaging with various undersampling schemes. Results: The results of our experiments demonstrate that priors trained on complex images outperform priors trained only on magnitude images. Additionally, a prior trained on a larger dataset exhibits higher robustness. Finally, we show that the generative priors are superior to L1 -wavelet regularization for compressed sensing parallel imaging with high undersampling. Conclusion: These findings stress the importance of incorporating phase information and leveraging large datasets to raise the performance and reliability of the generative priors for MRI reconstruction. Phase augmentation makes it possible to use existing image databases for training.

Generative Priors for MRI Reconstruction Trained from Magnitude-Only Images Using Phase Augmentation

TL;DR

This work addresses accelerating MRI reconstruction by learning priors from magnitude-only images. It introduces phase augmentation via a diffusion-based prior to synthesize realistic complex-valued images, enabling six priors trained on varying dataset sizes and complex/magnitude data to regularize linear and nonlinear reconstructions. The study demonstrates that priors trained on complex-valued data outperform magnitude-only priors, with larger datasets offering greater robustness; diffusion priors further improve 3D reconstructions and often exceed conventional -wavelet regularization under high undersampling. Practically, phase augmentation provides a scalable workflow to leverage existing magnitude-only image databases for robust, phase-informed MRI reconstruction across diverse undersampling schemes and coil configurations.

Abstract

Purpose: In this work, we present a workflow to construct generic and robust generative image priors from magnitude-only images. The priors can then be used for regularization in reconstruction to improve image quality. Methods: The workflow begins with the preparation of training datasets from magnitude-only MR images. This dataset is then augmented with phase information and used to train generative priors of complex images. Finally, trained priors are evaluated using both linear and nonlinear reconstruction for compressed sensing parallel imaging with various undersampling schemes. Results: The results of our experiments demonstrate that priors trained on complex images outperform priors trained only on magnitude images. Additionally, a prior trained on a larger dataset exhibits higher robustness. Finally, we show that the generative priors are superior to L1 -wavelet regularization for compressed sensing parallel imaging with high undersampling. Conclusion: These findings stress the importance of incorporating phase information and leveraging large datasets to raise the performance and reliability of the generative priors for MRI reconstruction. Phase augmentation makes it possible to use existing image databases for training.
Paper Structure (17 sections, 9 equations, 9 figures, 1 table, 1 algorithm)

This paper contains 17 sections, 9 equations, 9 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Theory
  3. Linear and nonlinear reconstruction
  4. Learned priors as regularization
  5. Methods
  6. Preprocessing of the training dataset
  7. Phase augmentation
  8. Training of priors
  9. Experimental evaluation
  10. Results
  11. The influence of phase maps
  12. The influence of the size of dataset
  13. 3D reconstruction using diffusion priors
  14. Discussion
  15. Conclusion
  16. ...and 2 more sections

Figures (9)

  • Figure 1: The proposed workflow for extracting prior knowledge and using it for regularization in image reconstruction. It comprises data preparation, phase augmentation, generative modeling, and concludes with the use as learned regularizers in reconstruction.
  • Figure 2: Human brain images. On the left, the original magnitude-only image is compared to the magnitude of a corresponding image generated using phase augmentation. On the right, the phase maps of three different generated images are shown.
  • Figure 3: Comparison of images reconstructed using PICS using the priors PSM, PLM, PSC, PLC, DSC in comparison to an $\ell_1$-wavelet reconstruction and a reference (c.f. error maps in the supplementary). The top two rows (1D) present the results for 5-fold acceleration along phase-encoding direction with 30 calibration lines. The bottom two rows (Poisson) show the results using a Poisson-disc acquisition of 8.2x-undersampling. PSNR and SSIM values are shown in white text.
  • Figure 4: Comparison of images reconstructed using NLINV using the priors PSM, PLM, PSC, PLC, DSC in comparison to an $\ell_1$-wavelet reconstruction and a reference (c.f. error maps in the supplementary). The top two rows (1D) present the results for 5-fold acceleration along phase-encoding direction with 30 calibration lines. The bottom two rows (Poisson) show the results using a Poisson-disc acquisition of 8.2x-undersampling. PSNR and SSIM values are shown in white text.
  • Figure 5: Comparison of images reconstructed using NLINV and PICS using the priors PSM, PLM, PSC, PLC, DSC for a 2 $\times$ 3 sampling pattern in comparison to an $\ell_1$-wavelet reconstruction and a reference (c.f. error maps in the supplementary). PSNR and SSIM values are shown in white text. Artifacts (red arrow) are introduced by the priors trained on magnitude images when using PICS.
  • ...and 4 more figures