Efficient calculation of g-factors for CG-SENSE in high dimensions: noise amplification in random undersampling

Mehmet Akçakaya, Tamer Basha, Warren J. Manning, Reza NezafatJournal of Cardiovascular Magnetic Resonance 2014 · DOI

Abstract

Parallel imaging techniques such as SENSE are essential for accelerating MRI acquisitions. A critical metric in these techniques is the g-factor, which quantifies noise amplification resulting from undersampling. While analytical calculation of g-factors is straightforward for uniform undersampling, it becomes computationally infeasible for arbitrary trajectories in high dimensions, such as those used in compressed sensing. This work introduces an efficient Monte-Carlo (MC) method for calculating g-factors in Conjugate Gradient SENSE (CG-SENSE) and applies it to analyze random Cartesian undersampling.

Introduction

Parallel imaging (PI) accelerates MRI by skipping k-space data points, using spatial information from multiple receiver coils to reconstruct the full image. However, this process inherently amplifies noise, a phenomenon characterized by the geometry factor, or g-factor. For standard SENSE with uniform undersampling, g-factors are easily computed. In contrast, modern 3D imaging often employs random undersampling trajectories to facilitate compressed sensing. For these complex trajectories, calculating g-factors analytically requires inverting a massive system matrix ($E^*E$), which is computationally prohibitive for high-resolution volumes. Consequently, researchers often lack precise knowledge of the noise characteristics introduced by their sampling patterns.

Methods

To overcome the computational bottleneck of matrix inversion, we utilized a Monte-Carlo (MC) approach. The method involves:

  1. Noise Simulation: Adding multiple instances of synthetic white noise to the system model.
  2. Reconstruction: Processing these noise-only datasets through a CG-SENSE reconstruction algorithm.
  3. Statistical Estimation: Calculating the standard deviation of the reconstructed noise images across multiple simulations to empirically determine the g-factor at each voxel.

Efficiency was improved by calculating only the diagonal elements of the noise covariance matrix. The method was first validated against an analytical solution in a 2D numerical simulation to determine the necessary number of MC iterations for convergence. Subsequently, it was applied to 3D whole-heart imaging (1.3 mm isotropic resolution) using a 32-channel coil array, comparing 4-fold uniform and random undersampling.

Results and Discussion

Validation results demonstrated that the MC method accurately approximates the analytical g-factor. Convergence was observed to occur at approximately 50 simulations, making the approach computationally practical for high-dimensional data.

Validation of the Monte-Carlo method against an analytical solution. Convergence is achieved around 50 simulations (#MC=50).

In the 3D whole-heart experiments, a significant difference was observed between uniform and random undersampling. While uniform SENSE produces localized noise "folding" artifacts, random undersampling results in a more homogeneous noise distribution. Furthermore, the g-factor in random undersampling exhibited "semi-convergence," where the noise amplification level depends on the number of iterations in the CG-SENSE algorithm.

Comparison of g-factor maps for uniform and random undersampling across different CG-SENSE iterations. Random undersampling shows more uniform noise distribution.

Quantitatively, random undersampling yielded lower g-factors at high acceleration rates compared to traditional uniform patterns. For instance, near the ascending aorta, the random undersampling g-factor was lower in early iterations than the convergent g-factor of uniform SENSE, suggesting superior noise performance for practical reconstruction settings.

Conclusion

The proposed Monte-Carlo method provides an efficient and accurate means of calculating g-factors for complex, high-dimensional MRI trajectories. Our findings suggest that random undersampling offers distinct advantages in 3D parallel imaging by providing more uniform noise amplification and potentially lower overall noise levels at high acceleration factors. This tool enables the optimization of sampling patterns for advanced cardiac MRI applications.

Summary updated

Abstract

Background SENSE [1,2] is one of the most used parallel imaging techniques. In [1], uniform undersampling was employed to efficiently reconstruct an unalised image, whereas in [2], a conjugate gradient-based method (CG-SENSE) was used for reconstruction with arbitrary trajectories. SENSE framework allows the calculation of g-factors, characterizing the noise amplification for a given k-space trajectory and coil configuration [1]. However, calculation of g-factors for arbitrary trajectories in high dimensions is time-consuming [3]. Furthermore, noise characteristics of random undersampling, used in compressed sensing, is not wellunderstood. In this work, we use a Monte-Carlo (MC) method for fast calculation of g-factors for CG-SENSE similar to [4,5] and apply it to random Cartesian undersampling trajectories. Theory: SENSE involves a prewhitening step [1,2], thus without loss of generality, we assume white noise. SENSE reconstruction solves minm ||y Em||2, where E is the system matrix, and y are the undersampled measurements. The g-factor for the k voxel is given by gk = √([E*E] -1 k,k [E*E]k,k). Inverting E*E is not feasible in high dimensions. Instead we note the gk corresponds to the k diagonal of the reconstruction noise covariance matrix (for normalized coil sensitivities), where nrecon = (E*E) E*nmeas, and nmeas is measurement noise with identity covariance matrix. We calculate the sample correlation matrix using a MC approach (since sample mean goes to 0), as 1/(p-1)∑p n p recon (n p recon)* for p instances of nrecon. Note we only calculate and store the diagonal elements of this matrix, significantly increasing efficiency. Methods The MC method was first verified in a numerical simulation, where the g-factor was explicitly calculated for a 2D coil configuration, to determine how many MC simulations suffice. Whole-heart imaging was performed with an isotropic resolution of 1.3 mm using a 32-channel coil array. Two 4-fold accelerated acquisitions were performed, one with uniform undersampling (2 × 2 in the ky-kz plane) and one with random undersampling. Coil sensitivity maps were exported. Images were reconstructed using SENSE (for uniform) and CG-SENSE (for both). g-factors were also calculated with the proposed approach.
BibTeX
@article{akcakaya2014efficient,
  author = {Mehmet Akçakaya and Tamer Basha and Warren J. Manning and Reza Nezafat},
  title = {{Efficient calculation of g-factors for CG-SENSE in high dimensions: noise amplification in random undersampling}},
  journal = {Journal of Cardiovascular Magnetic Resonance},
  year = {2014},
  doi = {10.1186/1532-429x-16-s1-w28},
}
Figure