Variable density compressed sensing in MRI. Theoretical vs heuristic sampling strategies

Nicolas Chauffert, Philippe Ciuciu, Pierre WeissISBI 2013 · DOI

TL;DR

This paper tackles the challenge of speeding up MRI scans by acquiring less data. They propose a novel two-stage sampling strategy that intelligently samples the "center" of the MRI data (k-space) more densely, then uses a theoretically optimal distribution for the rest. This approach significantly improves reconstruction quality compared to existing methods, making faster and more efficient MRI scans a reality.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but scan times can be long, leading to patient discomfort, motion artifacts, and higher costs. A primary way to reduce scan time is to acquire fewer data points in the k-space (the frequency domain representation of the MRI signal). Compressed Sensing (CS) theory offers a theoretical framework to reconstruct a high-quality image from these undersampled measurements, provided the image is "sparse" in a particular representation (like a wavelet basis).

However, a key challenge has been determining the optimal way to undersample the k-space – the "synthesis problem." While theoretical CS approaches suggest independent random sampling from a specific probability distribution, this often doesn't perform optimally in practice for MRI. This is because MRI images, particularly their low-frequency components, are not as sparse as CS theory often assumes, especially in the center of the k-space. This paper aims to bridge this gap by proposing a practical, yet theoretically grounded, sampling strategy that accounts for the specific sparsity characteristics of MRI data.

Methods

The authors introduce a two-stage sampling strategy for MRI k-space acquisition.

  1. Acquiring the k-space center: They first recognize that the low-frequency components of MRI images, which reside in the center of the k-space, are not very sparse. To accurately capture this information, they propose fully sampling this central region. This is a departure from purely theoretical CS approaches that would undersample even this region.
  2. Sampling high frequencies: For the remaining high-frequency components, which are more amenable to CS theory and sparsity assumptions, they apply a theoretically derived optimal sampling distribution. This distribution, denoted as $\pi$, is derived from minimizing an upper bound on the number of measurements required for exact signal recovery. They also introduce a refined distribution, $\pi^*$, which accounts for the rejection of already sampled points.

The reconstruction process then involves solving a $\ell_1$-minimization problem to recover the image from the acquired undersampled data. They compare their proposed method against various heuristic sampling strategies, including radial, spiral, and polynomial density sampling.

Datasets

The paper evaluates its proposed method using a reference 256x256 MRI image (shown in Fig. 1(a)). This image is used to demonstrate the reconstruction quality across different sampling schemes. No new datasets are introduced; rather, the focus is on evaluating sampling strategies on a typical anatomical MRI image.

Key Insights and Ideas

  • MRI data isn't uniformly sparse: The core insight is that the assumption of uniform sparsity across all frequency components of an MRI image (in a wavelet basis) is flawed. Low-frequency components are particularly problematic for standard CS.
  • Intelligent undersampling is key: Simply applying theoretical optimal sampling distributions derived for uniformly sparse signals doesn't translate to the best practical results for MRI. A more nuanced approach is needed.
  • The k-space center is special: Fully sampling the k-space center is crucial for capturing essential low-frequency information that is not well-represented by sparse models. This significantly boosts reconstruction quality.
  • Bridging theory and practice: The two-stage approach effectively combines theoretical guarantees for high-frequency components with a practical strategy for the low-frequency ones, leading to improved performance.
  • Theoretical bounds can be improved: The paper shows how accounting for sample rejection can refine theoretical bounds, making them more relevant.

Key Results

  • The proposed two-stage sampling strategy, especially when combined with the $\pi^*$ distribution for high frequencies, drastically improves reconstruction results compared to standard CS approaches and heuristic methods.
  • Sampling the entire k-space center in the first stage leads to a significant gain in PSNR (Peak Signal-to-Noise Ratio) compared to methods that undersample the entire k-space. For example, the gain for $\pi^*$ is reported as 2.49 dB.
  • The standard deviation of reconstruction results for random sampling schemes is strongly reduced using the two-stage approach, indicating more consistent and reliable reconstructions.
  • The two-stage sampling schemes perform better than spiral and radial patterns irrespective of whether angles are regular or random.
  • While the proposed method achieves results comparable to state-of-the-art polynomial density sampling, it is argued to be more aligned with the sparsity assumptions required for CS design.

Limitations and Future Work

  • Theoretical bounds might not be optimal: The authors acknowledge that the theoretical bounds derived might not be perfectly optimal. Minimizing these bounds might not yield the absolute best sampling distribution in practice, especially considering that sparsity can vary across different wavelet subbands.
  • Continuity of trajectories: For practical MRI implementation, continuous sampling trajectories are often preferred. The current method generates discrete sampling points. The authors are investigating the design of optimal continuous trajectories by combining optimal drawing with continuity using Markov chains.
  • Understanding performance differences: The exact reasons why the $\pi^*$ distribution sometimes yields slightly less accurate results than high-order polynomial densities are still under investigation, with hypotheses focusing on whether the signal remains sufficiently sparse after the first stage or if the theoretical bounds themselves need refinement.

Summary updated

BibTeX
@inproceedings{chauffert2013variable,
  author = {Nicolas Chauffert and Philippe Ciuciu and Pierre Weiss},
  title = {{Variable density compressed sensing in MRI. Theoretical vs heuristic sampling strategies}},
  booktitle = {ISBI},
  year = {2013},
  doi = {10.1109/isbi.2013.6556471},
}
Figure