Suppression of MRI Truncation Artifacts Using Total Variation Constrained Data Extrapolation

Kai Tobias Block, Martin Uecker, Jens FrahmInternational Journal of Biomedical Imaging 2008 · DOI

TL;DR

This paper introduces a novel method to reduce "Gibbs ringing" artifacts in MRI images caused by incomplete k-space sampling. By assuming objects are piecewise constant, they extrapolate missing k-space data using total variation minimization, effectively suppressing artifacts without sacrificing image resolution. A denoising extension is also presented for noisy acquisitions.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) reconstructs images by transforming sampled data from "k-space" (which encodes spatial frequencies) into the image domain. A fundamental limitation is that k-space is never fully sampled due to practical acquisition constraints. This incomplete sampling leads to signal truncation, which manifests as "Gibbs ringing" artifacts – spurious oscillations around sharp edges in the reconstructed image. These artifacts are particularly noticeable in low-resolution scans, which are often used for faster acquisition. Traditional methods to mitigate Gibbs ringing, like low-pass filtering, effectively blur the image, leading to a loss of valuable spatial detail. This paper aims to address this trade-off by developing a method that reduces Gibbs ringing without compromising image resolution.

Methods

The core of the proposed method relies on the assumption that the underlying object being imaged is piecewise constant. This assumption is then translated into a mathematical optimization problem: minimizing the "total variation" (TV) of the reconstructed image. Total variation is a measure of the image's edge content, and minimizing it encourages images with large, uniform regions and sharp transitions, aligning with the piecewise-constant assumption.

Specifically, the method extrapolates the measured k-space data by adding synthetic frequencies. This extrapolation is guided by a total variation minimization objective function. A nonlinear optimization algorithm (CG-Descent) is used to solve this problem.

For situations where the MRI data is also contaminated with noise, an extended approach is introduced. This "edge-preserving denoising" method allows for slight deviations from the measured data while still minimizing TV. This dual objective helps to both suppress ringing artifacts and smooth out noise without blurring important edges.

The authors also discuss handling phase variations and multi-coil data, which can complicate the application of TV constraints. Their proof-of-principle approach involves a preprocessing step to combine coil data and obtain real-valued images.

Datasets

  • Shepp-Logan Phantom: A standard numerical phantom used for simulating MRI data.
  • Experimental Phantom: A physical phantom used for testing the method.
  • Human Brain In Vivo: Experimental data acquired from human subjects.

The paper does not introduce new benchmark datasets but rather uses established phantoms and real-world MRI data for evaluation.

Key Insights and Ideas

  • Piecewise-Constant Assumption as a Powerful Prior: The paper demonstrates that even a simple, unspecific prior like "piecewise-constant object" can be highly effective for k-space extrapolation and artifact suppression. This is a significant departure from methods that rely on more complex models or simply filter the data.
  • Total Variation Minimization for Artifact Reduction: Leveraging TV minimization not only for denoising but also for extrapolating missing k-space data is a key innovation. It elegantly links image properties to k-space reconstruction.
  • Resolution Preservation: The method's ability to reduce ringing artifacts without introducing blurring is a major advantage over traditional filtering techniques. In some cases, it can even lead to a mild resolution enhancement by sharpening the point-spread function.
  • Implicit Filtering in Extrapolation: The authors note that the extrapolation process implicitly filters the extrapolated data, preventing the introduction of new ringing artifacts at the extended k-space border. This is a desirable feature that helps to keep the method robust.

Key Results

  • Significant Reduction of Gibbs Ringing: Both simulations and experimental data show a substantial reduction in truncation artifacts (Gibbs ringing) compared to zero-padding.
  • Resolution Preservation: The TV-based reconstructions maintain image resolution, unlike filtered reconstructions which exhibit blurring.
  • Effective Denoising: The extended TV approach successfully reduces noise while preserving edges, demonstrating its utility for noisy acquisitions.
  • Comparison to Existing Methods: The proposed method is shown to be competitive with or superior to existing extrapolation techniques like the one by Constable and Henkelman, particularly in its robustness to parameter selection.
  • Computational Load: While effective, the current implementation requires a significant number of iterations (e.g., 3000), taking a few minutes on a standard microprocessor.

Limitations and Future Work

  • Piecewise-Constant Assumption: The method's performance is dependent on how well the true object conforms to the piecewise-constant assumption. Real-world objects are rarely perfectly piecewise constant, and deviations (e.g., flow artifacts) can limit the assumption's validity.
  • Computational Cost: The current implementation's high computational load is a limitation for routine clinical applications. The authors suggest that dedicated TV optimization algorithms could significantly improve processing speed.
  • Parameter Tuning: While less sensitive than some other methods, the denoising extension still requires manual adjustment of a weighting factor ($\lambda$) for optimal results.
  • Future Work: The authors suggest that developing more sophisticated methods to handle phase variations and multi-coil configurations, especially in conjunction with advanced techniques like parallel imaging, would be beneficial for routine clinical use. They also highlight the need for dedicated optimization algorithms to reduce computational time.

Summary updated

Abstract

The finite sampling of k-space in MRI causes spurious image artifacts, known as Gibbs ringing, which result from signal truncation at the border of k-space. The effect is especially visible for acquisitions at low resolution and commonly reduced by filtering at the expense of image blurring. The present work demonstrates that the simple assumption of a piecewise-constant object can be exploited to extrapolate the data in k-space beyond the measured part. The method allows for a significant reduction of truncation artifacts without compromising resolution. The assumption translates into a total variation minimization problem, which can be solved with a nonlinear optimization algorithm. In the presence of substantial noise, a modified approach offers edge-preserving denoising by allowing for slight deviations from the measured data in addition to supplementing data. The effectiveness of these methods is demonstrated with simulations as well as experimental data for a phantom and human brain in vivo.
BibTeX
@article{block2008suppression,
  author = {Kai Tobias Block and Martin Uecker and Jens Frahm},
  title = {{Suppression of MRI Truncation Artifacts Using Total Variation Constrained Data Extrapolation}},
  journal = {International Journal of Biomedical Imaging},
  year = {2008},
  doi = {10.1155/2008/184123},
}
Figure