[Paper Review] Reducing the Dimensionality of Optimal Experiment Design for Magnetic Resonance Fingerprinting

Nikolai J. Mickevicius, Andrew S. Nencka+1 more|arXiv (Cornell University)|Oct 1, 2020

Advanced MRI Techniques and Applications13 references4 citations

TL;DR

This paper proposes a low-dimensional parametric model for optimal experiment design in magnetic resonance fingerprinting (MRF), reducing the number of free parameters from thousands to just 8 by representing flip angle patterns as weighted sums of overlapping Gaussian functions. The method achieves T₂ precision comparable to full optimization in under one minute, significantly accelerating computation while maintaining high accuracy.

ABSTRACT

Nuclear magnetic resonance signal dynamics as described by the Bloch equations are highly complex and often are without closed form solutions. This is especially the case for quantitative magnetic resonance fingerprinting (MRF) scans in which acquisition parameters are varied to efficiently probe the parameter space. As previously demonstrated, optimization of the pattern in which the MRI acquisition parameters are varied relative to the variance in the target quantitative parameters can improve experiment design. This process, however, relies on large scale non-linear optimizations with hundreds to thousands of unknowns. As such, the numerical optimization is extremely time consuming, highly sensitive to guesses, and prone to getting caught in local minima. Here, we describe a method to reduce the complexity of the optimal MRF experiment design by constraining the solutions to span a predetermined low-dimensional subspace. Compared with standard flip angle patterns in calibration phantom experiments, precision in T2 can be increased by optimizing as few as 8 coefficients in less than one minute of computation time.

Motivation & Objective

• Address the computational intractability of full-scale non-linear optimization in MRF experiment design, which can take hours and is sensitive to initial guesses.
• Reduce the dimensionality of the optimization space for MRF acquisition parameters (e.g., flip angles and repetition times) to make global optimization feasible.
• Enable rapid, robust, and scanner-specific MRF protocol design by constraining solutions to a low-dimensional subspace.
• Maintain or improve estimation precision of quantitative MRI parameters (T₁, T₂, M₀) despite reduced parameter space.
• Demonstrate that a small number of basis function coefficients can yield performance comparable to full optimization.

Proposed method

• Represent the flip angle pattern as a linear combination of partially overlapping Gaussian basis functions, reducing the number of unknowns from hundreds to tens.
• Optimize the coefficients of these basis functions using the Cramér-Rao bound (CRB) to minimize the variance in unbiased estimation of T₁, T₂, and M₀.
• Use extended phase graph (EPG) formalism to compute signal evolution and its derivatives with respect to tissue parameters, enabling accurate Fisher information matrix calculation.
• Compute partial derivatives of the signal evolution with respect to T₁, T₂, and M₀ using recursive formulations based on EPG equations.
• Construct the Fisher information matrix from the inner products of the derivative vectors and invert it to compute the Cramér-Rao lower bound (rCRB).
• Perform numerical optimization over the small set of basis function coefficients to minimize the rCRB for T₂ estimation.

Experimental results

Research questions

• RQ1 Can a low-dimensional parametric representation of MRF flip angle patterns achieve comparable estimation precision to full-scale optimization?
• RQ2 How many basis function coefficients are needed to achieve near-optimal performance in T₂ estimation?
• RQ3 Can the optimization time be reduced from hours to under a minute without sacrificing estimation accuracy?
• RQ4 Does the low-dimensional approach remain robust to initial parameter guesses and avoid local minima?
• RQ5 Can this framework be adapted for scanner-specific corrections (e.g., B₁⁺, ΔB₀, eddy currents) due to its reduced complexity?

Key findings

• Optimizing just 8 coefficients in the Gaussian basis representation reduced computation time to less than one minute while achieving T₂ estimation precision comparable to full optimization.
• The low-dimensional method significantly outperformed standard flip angle patterns in calibration phantom experiments, improving T₂ precision.
• The method demonstrated robustness to initial guesses and avoided local minima, unlike full-scale optimizations that are highly sensitive to initialization.
• The Cramér-Rao lower bound (rCRB) was effectively minimized using the parametric basis, confirming improved theoretical estimation precision.
• The approach enables rapid, iterative, and scanner-specific MRF protocol tuning, making it practical for clinical deployment.
• The use of overlapping Gaussian functions provided smooth, physically plausible flip angle patterns that enhanced signal diversity and parameter sensitivity.