Phase reconstruction from velocity-encoded MRI measurements – A survey of sparsity-promoting variational approaches

Martin Benning, Lynn F. Gladden, Daniel J. Holland, Carola‐Bibiane Schönlieb, Tuomo ValkonenJournal of Magnetic Resonance 2013 · DOI

TL;DR

This paper surveys and compares different sparsity-promoting variational approaches for reconstructing velocity-encoded MRI images. The authors introduce a robust guideline for setting up these reconstructions, particularly highlighting the benefits of Bregman iteration for contrast enhancement and the Morozov discrepancy principle for determining the optimal stopping point, even with limited prior knowledge of noise levels.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a powerful tool for visualizing fluid flow (e.g., blood in vessels, industrial fluid dynamics). However, acquiring velocity-encoded MRI data is notoriously time-consuming. To speed this up, researchers often undersample the data in k-space (the frequency domain representation of the image). This undersampling, combined with inherent noise, makes direct reconstruction difficult and leads to artifacts. While various reconstruction methods exist, choosing the right approach, especially one that can handle both smooth and sharp features common in flow images, remains a challenge. This paper aims to provide a systematic survey and practical guidance for selecting and implementing effective reconstruction techniques for velocity-encoded MRI.

Methods

The core of the paper revolves around variational regularization methods for solving the ill-posed inverse problem of reconstructing an image from undersampled k-space data. The general formulation involves minimizing a regularization functional $J(u)$ subject to a fidelity term that measures how well the reconstructed image $u$ matches the acquired data $f$.

The authors explore three main regularization functionals:

  • Wavelets (WAV): Penalizes the L1-norm of the wavelet transform of the image, promoting sparsity in the wavelet domain.
  • Total Variation (TV): Penalizes the L1-norm of the image gradient, effective at preserving edges but can lead to "stair-casing" artifacts.
  • Total Generalized Variation (TGV²): A higher-order regularization that balances first and second-order derivatives, aiming to overcome TV's limitations by better handling smooth variations while still preserving sharp features.

To address the contrast loss often seen in standard variational methods, they employ Bregman iteration. This iterative approach effectively enhances contrast by repeatedly solving a similar minimization problem, adding back the residual to the data at each step.

A crucial aspect is determining when to stop the Bregman iterations. The authors advocate for Morozov's discrepancy principle, which stops iterations when the residual error matches the expected noise level. This provides an objective criterion for selecting the regularization parameter $\alpha$.

Datasets

The paper uses two types of datasets:

  1. Computer-generated data: A simulated velocity field of water flowing through a packed bed of spheres, created using a lattice-Boltzmann simulation. This dataset is used extensively for evaluating different methods and parameters.
  2. Real data: Velocity-encoded MRI data of water flowing through a packed bed, acquired using a standard MRI scanner. This dataset is used to validate the findings from the computer-generated data.

Key Insights and Ideas

  • Bregman Iteration for Contrast: Bregman iteration is crucial for recovering the lost contrast in undersampled MRI reconstructions, significantly improving visual quality without amplifying noise, especially for L1-norm based regularizers.
  • Discrepancy Principle as a Robust Stopping Criterion: Morozov's discrepancy principle provides a reliable and objective way to stop Bregman iterations, especially when the noise level is known or can be estimated. This avoids manual tuning and leads to consistent results.
  • TGV² Outperforms TV and WAV for Phase: For velocity-encoded MRI, where phase information is critical, Total Generalized Variation (TGV²) generally offers superior reconstruction quality compared to Total Variation (TV) and Wavelets (WAV), particularly in preserving smooth phase variations and avoiding stair-casing artifacts.
  • Sampling Strategy Matters: While random sampling is generally preferred in compressed sensing, the paper shows that specific patterns (like Gaussian or spiral) can perform better than others (like line sampling) for velocity MRI, especially when undersampling is aggressive.
  • Guideline for Practical Reconstruction: The paper provides a clear, step-by-step guideline for practitioners to set up and perform variational reconstructions for velocity-encoded MRI, covering regularization choice, parameter selection, and stopping criteria.

Key Results

  • Quantitative Improvements: Tables and figures demonstrate that Bregman iteration with TGV² regularization, combined with the discrepancy principle, achieves higher PSNR and SSIM values for both magnitude and phase compared to standard TV or WAV, and backprojection, across various sampling patterns and noise levels. For instance, SSIM values for phase can reach above 0.95 with TGV².
  • Qualitative Improvements: Visual comparisons (Figures 8, 9, 10, 11, 12, 14) clearly show that TGV² reconstructions exhibit better contrast and fewer artifacts (like stair-casing from TV or blockiness from WAV) than other methods, especially in the phase images which are critical for velocity information.
  • Parameter Selection: The paper provides practical advice on selecting the regularization parameter $\alpha$ and $\beta$, suggesting that $\alpha \sim c ||F^{-1}S^T y||_\infty$ can be a good starting point for over-regularization, and that the discrepancy principle helps fine-tune this.

Limitations and Future Work

  • Regularizer Choice: While TGV² is shown to be superior for phase reconstruction in this context, the authors suggest that a combination of TV for magnitude and TGV² for phase might offer the best overall performance for piecewise-constant magnitude images with smooth phase.
  • Sampling Schemes: The paper focuses on a selection of sampling schemes. Further investigation into more complex or adaptive sampling strategies could yield additional improvements.
  • Real Data Validation: While a real data set was used, the authors acknowledge that for real data, ground truth is often unavailable, making direct quantitative comparison challenging. Future work could involve more extensive validation on diverse real-world scenarios.
  • Computational Cost: While Bregman iteration and higher-order regularizers like TGV² offer better quality, they can be computationally more intensive. Optimizing these algorithms for speed remains an ongoing area of research.

Summary updated

BibTeX
@article{benning2013phase,
  author = {Martin Benning and Lynn F. Gladden and Daniel J. Holland and Carola‐Bibiane Schönlieb and Tuomo Valkonen},
  title = {{Phase reconstruction from velocity-encoded MRI measurements – A survey of sparsity-promoting variational approaches}},
  journal = {Journal of Magnetic Resonance},
  year = {2013},
  doi = {10.1016/j.jmr.2013.10.003},
}
Figure