SLfRank: Shinnar-Le-Roux Pulse Design with Reduced Energy and Accurate Phase Profiles using Rank Factorization

Frank Ong, Michael Lustig, Shreyas Vasanawala, John M. Pauly, Vasanawala, Shreyas, Pauly, John

TL;DR

This paper introduces SLfRank, a novel algorithm that improves the design of radio-frequency (RF) pulses for MRI. By reformulating the Shinnar-Le-Roux (SLR) algorithm using convex relaxation and rank factorization, SLfRank generates pulses with significantly lower energy and more accurate phase profiles, potentially leading to faster and more efficient MRI scans.

What Problem Does This Paper Address?

Magnetic Resonance Imaging (MRI) relies on precisely timed radio-frequency (RF) pulses to excite specific tissues and generate images. The Shinnar-Le-Roux (SLR) algorithm is a widely used method for designing these RF pulses, particularly for achieving large "flip angles" (the degree to which magnetization is tipped). However, the standard SLR algorithm has a limitation: it solves for two key components of the pulse design sequentially, rather than jointly. This sequential approach can lead to suboptimal pulses that have higher energy (which is directly related to the Specific Absorption Rate or SAR, a safety concern) and less accurate phase profiles, especially for applications like slice excitation where precise phase control is crucial. The paper aims to overcome this bi-linear coupling issue in the SLR design process to achieve better pulse performance.

Methods

The core innovation of SLfRank lies in its approach to the bi-linear coupling inherent in the SLR algorithm. The original SLR algorithm maps RF pulses to pairs of polynomials (Cayley-Klein parameters, denoted as 'a' and 'b') and then designs these polynomials to match desired magnetization profiles. The challenge is that the relationship between these polynomials and the actual magnetization is bi-linear.

SLfRank tackles this by:

  • Convex Relaxation: Instead of directly solving the bi-linear problem, SLfRank uses a technique borrowed from low-rank matrix recovery. It reformulates the problem as a convex optimization problem by relaxing the rank constraint of a key matrix (representing the pulse's underlying quaternion) into a positive semi-definite constraint. This allows for efficient and guaranteed convergence to a global optimum.
  • Joint Optimization: The convex formulation enables SLfRank to jointly optimize both CK polynomials ('a' and 'b') simultaneously. This is a significant departure from the sequential approach of the original SLR.
  • Rank Factorization: The name "SLfRank" comes from this use of rank factorization and convex relaxation. The algorithm aims to find a rank-one matrix solution, which corresponds to the ideal case, and the convex relaxation provides a way to approach this.
  • Direct Constraint Specification: The new method allows for direct specification of constraints on magnetization profiles, rather than indirect conversion through the CK parameters.

Datasets

Not discussed in the paper. The paper focuses on the algorithm's design and numerical validation using simulated pulse designs and one phantom experiment.

Key Insights and Ideas

  • Joint Optimization is Key: The most significant insight is that jointly optimizing the CK polynomials ('a' and 'b') is crucial for achieving both lower pulse energy and more accurate phase profiles, directly addressing the limitations of the sequential SLR approach.
  • Convex Relaxation as a Powerful Tool: The successful application of convex relaxation, a technique common in matrix recovery, to the bi-linear problem in RF pulse design demonstrates its versatility and effectiveness in solving complex optimization problems in physics and engineering.
  • Energy Reduction is Significant: The paper consistently shows substantial reductions in pulse energy across various pulse types, which has direct implications for SAR reduction and potentially faster imaging.
  • Improved Phase Accuracy: For applications like slice excitation, SLfRank produces demonstrably more linear phase profiles, leading to better control over the excited slice.
  • Rank-One Solution as a Target: The algorithm's goal is to find a rank-one solution for the matrix P, which signifies an optimal pulse. The convex relaxation provides a path to this ideal solution, and the paper shows that the relaxation is often tight in practice.

Key Results

SLfRank consistently outperforms the original SLR algorithm in terms of pulse energy and phase profile accuracy across various pulse types (excitation, inversion, saturation, spin-echo refocusing).

  • Energy Reduction: Reductions in pulse energy range from 5.40% (maximum phase saturation) to 37.8% (minimum phase excitation with TBW=4, n=64). For linear phase excitation pulses with TBW=10, energy is reduced by 18.1%.
  • Peak Amplitude Reduction: Peak pulse amplitude is also reduced in most cases, for example, by 9.1% for linear phase excitation pulses.
  • Phase Profile Accuracy: For linear phase excitation, SLfRank produces a "much flatter phase response" and more linear phase profiles compared to SLR.
  • Phantom Experiment: A phantom experiment demonstrated that SLfRank-designed pulses produce slice profiles that match numerical simulations well, with a flatter phase profile than SLR.
  • Multi-band Imaging: A multi-band RF pulse designed with SLfRank showed a 19% reduction in RF energy compared to an SLR-designed pulse, with similar image quality.

Limitations and Future Work

  • Computational Cost: SLfRank is computationally more expensive than the original SLR algorithm, taking around two minutes to run compared to less than a second for SLR on the tested hardware. The authors suggest leveraging fast Fourier transforms, GPUs, and using SLR pulses or coarser resolution SLfRank designs for initialization to improve computation time.
  • Minimum Phase Excitation Gap: For minimum phase excitation pulses, there is a slight gap between the convex program's solution and the rank-constrained problem, and it's unclear if this is a fundamental limitation or an artifact of the relaxation.
  • Initial Magnetization Assumption: Both SLR and SLfRank assume initial magnetization starts from equilibrium, except for spin-echo refocusing. Generalizing the algorithm to arbitrary starting magnetization is a direction for future work.
  • Root Flipping: The authors suggest exploring "root flipping" as a potential avenue to further reduce peak RF amplitude.
  • Exotic Phase Profiles: SLfRank's ability to design more exotic phase profiles opens up possibilities for applications beyond standard excitation, such as quadratic phase and multi-band designs.

Summary updated

Abstract

The Shinnar-Le-Roux (SLR) algorithm is widely used to design frequency selective pulses with large flip angles. We improve its design process to generate pulses with lower energy (by as much as 26%) and more accurate phase profiles. Concretely, the SLR algorithm consists of two steps: (1) an invertible transform between frequency selective pulses and polynomial pairs that represent Cayley-Klein (CK) parameters and (2) the design of the CK polynomial pair to match the desired magnetization profiles. Because the CK polynomial pair is bi-linearly coupled, the original algorithm sequentially solves for each polynomial instead of jointly. This results in sub-optimal pulses. Instead, we leverage a convex relaxation technique, commonly used for low rank matrix recovery, to address the bi-linearity. Our numerical experiments show that the resulting pulses are almost always globally optimal in practice. For slice excitation, the proposed algorithm results in more accurate linear phase profiles. And in general the improved pulses have lower energy than the original SLR pulses.
BibTeX
@article{ong2021slfrank,
  author = {Frank Ong and Michael Lustig and Shreyas Vasanawala and John M. Pauly and Vasanawala, Shreyas and Pauly, John},
  title = {{SLfRank: Shinnar-Le-Roux Pulse Design with Reduced Energy and Accurate Phase Profiles using Rank Factorization}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2021},
}
Figure