PaperlayerView exact revision on arXiv

Source-linked AI summary

Optimal Experiment Design for Magnetic Resonance Fingerprinting: Cram\'er-Rao Bound Meets Spin Dynamics

Bo Zhao, Justin P. Haldar, Congyu Liao, Dan Ma, Yun Jiang, Mark A. Griswold, Kawin Setsompop, Lawrence L. Wald

arXiv:1710.08062v8eess.SP

TL;DR

MR fingerprinting experiment design lacks finite-duration quality guidance, and T2 maps can be less accurate than T1 maps for short acquisitions. This paper models spin dynamics and uses the Cramér-Rao bound to optimize acquisition parameters under physical and reconstruction constraints. Optimized experiments improve T2-map accuracy while maintaining similar or slightly better T1-map accuracy, and their acquisition parameters appear structured.

  • Problem

    Prior theoretical analysis did not guide the quality or design of finite-duration MR fingerprinting experiments, and T2-map accuracy can be poor for short acquisitions.

    Source: §I · p. 1

  • Method

    The framework models spin dynamics as a discrete-time dynamic system and uses the Cramér-Rao bound to optimize acquisition parameters for SNR efficiency under physical and reconstruction constraints.

    Source: §I · p. 2

  • Results

    Optimized experiments substantially improved T2-map accuracy while yielding similar or slightly better T1-map accuracy; T2 standard deviation fell by about a factor of two.

    Source: §IV · pp. 14–15, §V · p. 16

  • Takeaways & Limitations

    Optimized acquisition parameters appeared highly structured rather than randomly or pseudo-randomly varying.

    Source: §V · p. 16

    In vivo bias assessment is complicated by the lack of ground truth and by model mismatches such as partial-volume effects.

    Source: §IV · p. 15

Abstract

from arXiv · show

Magnetic resonance (MR) fingerprinting is a new quantitative imaging paradigm, which simultaneously acquires multiple MR tissue parameter maps in a single experiment. In this paper, we present an estimation-theoretic framework to perform experiment design for MR fingerprinting. Specifically, we describe a discrete-time dynamic system to model spin dynamics, and derive an estimation-theoretic bound, i.e., the Cramer-Rao bound (CRB), to characterize the signal-to-noise ratio (SNR) efficiency of an MR fingerprinting experiment. We then formulate an optimal experiment design problem, which determines a sequence of acquisition parameters to encode MR tissue parameters with the maximal SNR efficiency, while respecting the physical constraints and other constraints from the image decoding/reconstruction process. We evaluate the performance of the proposed approach with numerical simulations, phantom experiments, and in vivo experiments. We demonstrate that the optimized experiments substantially reduce data acquisition time and/or improve parameter estimation. For example, the optimized experiments achieve about a factor of two improvement in the accuracy of $T_2$ maps, while keeping similar or slightly better accuracy of $T_1$ maps. Finally, as a remarkable observation, we find that the sequence of optimized acquisition parameters appears to be highly structured rather than randomly/pseudo-randomly varying as is prescribed in the conventional MR fingerprinting experiments.

I. INTRODUCTION

The paper addresses gaps in the theory and practical performance of MR fingerprinting by developing an estimation-theoretic framework for experiment design. It models spin dynamics in discrete time and extends the model to voxel-level magnetization observations.

I. INTRODUCTION

  • Prior analysis did not guide the quality or design of finite-duration experiments, while T2-map accuracy can be poor for short acquisitions.The practical goal is to improve mapping quality without lengthening experiments, or shorten experiments without sacrificing mapping quality.Source: §I · pp. 1–2
  • The framework uses the Cramér-Rao bound to select acquisition parameters for maximal SNR efficiency under physical and reconstruction constraints.The paper also reports that optimized parameter sequences appear structured rather than random or pseudo-random.Source: §I · p. 2

A. Signal Model

  • A discrete-time state-space model represents magnetization at the finite time points of interest in a voxel, capturing the relevant spin dynamics.The model uses isochromat-specific state equations and sums their magnetizations to obtain voxel-level magnetization.Source: A. Signal Model · p. 2, 1) State-Space Model · p. 3

1) State-Space Model:

  • The model treats the isochromats in a voxel separately, then sums their transverse magnetizations to describe the voxel signal.The subvoxels share the same parameter set.Source: 1) State-Space Model · p. 3
  • The resulting state-space model can describe spin dynamics across MR fingerprinting sequences and is nonlinear and time-varying.The IR-FISP example uses tissue parameters T1, T2, and M0 alongside time-varying acquisition parameters.Source: 1) State-Space Model · p. 3

2) Example Sequence:

For the IR-FISP example, the model represents RF excitation, spin relaxation, and spin dephasing through its system, input, and output components.

2) Example Sequence:

  • RF excitation, spin relaxation, and spin dephasing are the three processes driving magnetization evolution in the IR-FISP model.Phase dispersion for an isochromat is dominated by spoiler gradients, though other factors can contribute.Source: 2) Example Sequence · p. 3
  • The input matrix models longitudinal magnetization recovery, while the output matrix models RF excitation and relaxation.The output extracts transverse magnetization; it excludes dephasing because the spoiler gradient follows the echo time.Source: 2) Example Sequence · pp. 3–4
  • The CRB is unchanged by transforming the coordinate system used for the Bloch-equation model.The paper notes polar coordinates as an alternative to Cartesian coordinates in the rotating frame.Source: 2) Example Sequence · p. 3

B. Cram´er-Rao Bound

The CRB analysis uses a simplified noisy signal model, computes parameter sensitivities through difference equations, and bounds the covariance of unbiased estimates.

B. Cram´er-Rao Bound

  • The simplified data model omits spatial encoding because modeling non-Cartesian Fourier encoding can be computationally expensive.Despite this simplification, the paper reports that designs based on the model can be effective for highly undersampled experiments.Source: B. Cram´er-Rao Bound · p. 4
  • The assumed observation noise is independent, identically distributed Gaussian noise with covariance σ^2I.The model corresponds to a single-voxel NMR experiment or, more precisely, a fully sampled Cartesian acquisition with discrete Fourier reconstruction.Source: B. Cram´er-Rao Bound · p. 4
  • For unbiased estimators, the CRB lower-bounds covariance, and its diagonal entries bound the variance of individual tissue-parameter estimates.The bound is asymptotically achievable by maximum-likelihood estimation under the stated regularity conditions.Source: B. Cram´er-Rao Bound · p. 4
  • The signal Jacobian is computed by iterating difference equations obtained by differentiating the observation and state equations with respect to each parameter.The recursion starts from the stated magnetization and sensitivity initial conditions.Source: B. Cram´er-Rao Bound · p. 4

C. Optimal Experiment Design

The section uses a Cramér-Rao bound (CRB) criterion to select feasible acquisition parameters for maximal SNR efficiency. For IR-FISP, it minimizes a weighted CRB trace and can constrain flip-angle changes to promote smoother magnetization for undersampled reconstruction.

C. Optimal Experiment Design

  • The CRB bounds the variance of unbiased estimates, characterizes experiment SNR efficiency, and can guide acquisition time and encoding design.It provides a principled way to assess potential reliability and optimize the encoding process.Source: C. Optimal Experiment Design · p. 4
  • Because the CRB depends on tissue parameters, the design assumes a specified acquisition length and uses representative tissues as a practical compromise.A design universally optimal across all possible parameters in the range is described as often infeasible.Source: C. Optimal Experiment Design · p. 5
  • For IR-FISP, the A-optimality criterion minimizes the weighted trace of the CRB matrix, representing total tissue-parameter variance.Weights address differing parameter scales and allow emphasis on parameters relevant to an application.Source: C. Optimal Experiment Design · p. 5
  • The design selects flip angles and repetition times while fixing RF phases and echo times, subject to feasible acquisition constraints.The design variables are u[n] = [αn, TRn]; constraints can account for physical considerations such as SAR and total acquisition time.Source: C. Optimal Experiment Design · p. 5
  • For highly undersampled reconstruction, bounds on consecutive flip-angle changes promote smoother magnetization evolutions that can improve reconstruction accuracy.The paper notes that smooth evolutions can help pattern matching, low-rank reconstruction, and statistical reconstruction.Source: C. Optimal Experiment Design · p. 5
  • The resulting formulations are nonlinear and nonconvex; the authors use sequential quadratic programming to seek local minima.A total-time constraint could require jointly choosing acquisition length and parameters, often producing a highly nonconvex or mixed-integer problem.Source: C. Optimal Experiment Design · p. 5

III. RESULTS

The results section presents representative evaluations using numerical simulations, phantom experiments, and in vivo experiments. Its simulation setup uses a numerical brain phantom with ground-truth tissue maps.

III. RESULTS

  • The section presents representative results from numerical simulations, phantom experiments, and in vivo experiments.The results are intended to illustrate the performance of the proposed approach.Source: §III · p. 6

A. Simulations

  • The numerical brain phantom uses Brainweb T1, T2, and M0 maps as ground truth, with a 300 × 300 mm2 field of view and a 256×256 matrix.The simulations use the IR-FISP sequence and assume a homogeneous transmit RF field.Source: A. Simulations · p. 6

1) General Setup:

The setup compares conventional and optimized acquisition schemes, with optimized parameters appearing highly structured and additional flip-angle constraints smoothing magnetization evolutions. It evaluates fully sampled and highly undersampled data using reconstruction and CRB measures.

1) General Setup:

  • Bloch simulations generate contrast-weighted images using 400 isochromats per voxel and compare conventional, Optimized-I, and Optimized-II schemes.The conventional scheme is shown with N = 1000, including flip-angle and repetition-time trains and white- and gray-matter magnetization evolutions.Source: 1) General Setup · p. 6, Figure 2 · p. 7
  • Evaluations include fully sampled Cartesian data and highly undersampled spiral data, using one interleaf per time point from a 48-interleaf trajectory.Complex white Gaussian noise is added to the undersampled data at a prespecified level.Source: 1) General Setup · p. 6
  • Maximum-likelihood reconstruction is used for both acquisition setups, and accuracy is assessed by overall map error and voxelwise relative error.For fully sampled Cartesian data, maximum-likelihood reconstruction is equivalent to direct Fourier reconstruction followed by dictionary pattern matching.Source: 1) General Setup · p. 6
  • The optimized designs use three representative tissues and a manually selected weighting matrix, while tissue values are chosen from neuroimaging-relevant ranges rather than the phantom.The latter choice is stated to avoid the inverse crime.Source: 1) General Setup · p. 6
  • Optimized acquisition parameters appear highly structured across tested acquisition lengths, unlike the conventional scheme.The optimized repetition times are described as binary; similar structure was observed for the other tested lengths.Source: 1) General Setup · p. 7
  • Optimized-I produces marked magnetization oscillations early in the sequence, while Optimized-II's flip-angle constraints significantly suppress them.The resulting Optimized-II evolutions are smoother, a property the authors expect to benefit reconstruction from highly undersampled data.Source: 1) General Setup · p. 7
  • CRBs are evaluated across acquisition lengths at SNR = 33 dB, with normalized CRBs used to display T1 and T2 on the same scale.The comparison includes the conventional scheme and both optimized schemes.Source: 1) General Setup · p. 7

3) Evaluation of CRB:

The CRB analysis identifies substantially more room to improve T2 than T1 at short acquisition lengths, and optimized schemes improve reconstructed parameter-map accuracy. In fully sampled experiments, their advantage persists across acquisition lengths and SNR levels, with T2 gains more pronounced in the N=400 comparison.

3) Evaluation of CRB:

  • At N=400, optimized schemes reduce T2 nCRB by about a factor of two, especially at short acquisition lengths.The conventional scheme needs much longer acquisitions than the optimized schemes to attain a good T2 nCRB; T1 nCRB approaches its asymptote rapidly for all three schemes.Source: 3) Evaluation of CRB · pp. 7–8, Figure 4 · p. 8

4) Evaluation of fully-sampled experiments:

  • At N=400, both optimized schemes improve fully sampled T1 and T2 map accuracy, with a larger improvement for T2.Optimized-I is slightly better than Optimized-II, consistent with the CRB prediction.Source: 3) Evaluation of CRB · p. 8, Figure 5 · p. 9
  • Optimized schemes reduce normalized standard deviation for both maps, and standard deviation dominates normalized RMSE across all three schemes.The reduction is more substantial for T2; normalized bias is much smaller than standard deviation.Source: Figure 6 · p. 9, 4) Evaluation of fully-sampled… · p. 10
  • Across acquisition lengths and SNR levels, both optimized schemes outperform the conventional scheme in fully sampled experiments.The evaluation covers N=300–800 at SNR=33 dB and SNR levels from 28–38 dB at N=400.Source: 4) Evaluation of fully-sampled… · p. 10

5) Evaluation of highly-undersampled experiments:

In highly undersampled experiments, Optimized-II outperforms both the conventional scheme and Optimized-I across acquisition lengths and SNR levels. Phantom experiments used a nine-tube physical phantom with varied T1 and T2 values.

5) Evaluation of highly-undersampled experiments:

  • For highly undersampled maps, Optimized-II reduces normalized standard deviation and RMSE for both T1 and T2 versus the conventional scheme.It also reduces bias relative to Optimized-I; the authors associate this with smoother magnetization evolutions under the flip-angle-variation constraint.Source: 5) Evaluation of highly-undersampled… · p. 10, Figure 10 · p. 11
  • Optimized-II outperforms the conventional scheme and Optimized-I in overall T1 and T2 errors across acquisition lengths and SNR levels.The comparisons cover acquisition lengths at SNR=33 dB and SNR levels at N=400.Source: 5) Evaluation of highly-undersampled… · p. 10

B. Phantom Experiments

  • The phantom comprised nine plastic tubes with different Gadolinium and Agar concentrations, yielding T1 and T2 values relevant to neuroimaging.Experiments used a 3T Siemens Tim Trio scanner with a 32-channel head array coil.Source: B. Phantom Experiments · p. 10

1) General evaluation:

The optimized schemes improve phantom and in vivo T2 estimation while maintaining similar T1 performance. The discussion reports structured optimized schedules and notes limits of the model and design scope.

C. In Vivo Experiments

  • In vivo, both optimized schemes improve T2 standard deviation while providing similar T1 standard deviation; Optimized-II also improves variance for both parameters versus Optimized-I.Optimized-II additionally reduces T1 reconstruction artifacts in white matter relative to Optimized-I.Source: C. In Vivo Experiments · p. 12, Figure 16 · p. 15, Figure 17 · p. 15

1) General evaluation:

  • Optimized-II improves phantom T2-map accuracy over the conventional scheme while maintaining similar T1 accuracy, and outperforms Optimized-I particularly for T1.The experiments used N = 400; acquisition times were 5.28 sec, 5.24 sec, and 5.22 sec for the conventional scheme, Optimized-I, and Optimized-II, respectively.Source: 1) General evaluation · p. 12, Figure 13 · p. 13

IV. DISCUSSION

  • The design uses a simplified data model and flip-angle-variation constraints; directly optimizing highly undersampled experiments would require a larger model of contrast and spatial encoding.That higher-dimensional CRB calculation would be substantially more computationally expensive.Source: §IV · p. 13

1) General evaluation:

  • A factor of two reduction in T2 standard deviation across all phantom tubes was achieved by Optimized-II, with similar T1 standard deviation.The authors note that this T2 variance reduction could potentially correspond to quadrupling averaging time or upgrading to a higher-field scanner.Source: 1) General evaluation · p. 12, Figure 15 · p. 14, §IV · p. 15

IV. DISCUSSION

  • The optimized acquisition parameters appear highly structured, with optimized repetition times appearing binary; the paper reports this Bang-Bang structure as a new observation for MR fingerprinting.The authors suggest optimal-control principles as a possible explanation, while noting that a control-theoretic characterization is beyond the paper's scope.Source: §IV · p. 14
  • SQP scales polynomially with the number of design parameters and avoids dynamic programming's curse of dimensionality, but generally finds less favorable local minima.The paper reports substantial practical improvement over conventional acquisition despite this trade-off.Source: §IV · p. 14
  • CRB optimization assumes unbiased estimation, but real estimation error also includes bias, which remains difficult to assess in vivo without ground truth and amid model mismatch.Partial-volume effects can produce different magnetization evolutions under different excitations and complicate comparisons.Source: §IV · p. 15
  • This paper studies one implementation; alternative criteria, acquisition parameters, tissue targets, and practical constraints may suit other applications.Expanding the acquisition-parameter search could improve performance but make optimization more computationally expensive.Source: §IV · p. 15

V. CONCLUSION

The paper presents an estimation-theoretic framework for optimizing MR fingerprinting acquisition parameters. Its optimized experiments improve T2 accuracy while maintaining similar or slightly better T1 accuracy, and yield highly structured acquisition parameters.

V. CONCLUSION

  • The framework optimizes MR fingerprinting acquisition parameters by maximizing SNR efficiency while incorporating constraints advantageous to reconstruction.The authors model spin dynamics and use the Cramér-Rao bound in the design formulation.Source: §V · p. 16
  • The optimized experiments substantially improve T2-map accuracy while providing similar or slightly better T1-map accuracy.Source: §V · p. 16
  • Optimized acquisition parameters appear highly structured rather than randomly or pseudo-randomly varying as in conventional MR fingerprinting.Source: §V · p. 16

VII. APPENDIX

The appendix illustrates the Jacobian-derivation procedure for the IR-FISP sequence. It describes calculating magnetization derivatives with respect to T1 by iteration from initial conditions.

VII. APPENDIX

  • The appendix derives the Jacobian Jn(θ) for IR-FISP to illustrate the procedure described in equations (13) and (14).For this sequence, θ contains T1, T2, and M0.Source: §VII · p. 16
  • For IR-FISP, the state-space model is summarized before derivatives of m[n] with respect to T1, T2, and M0 are calculated.Source: §VII · p. 16

A. Derivative of m[n] with respect to T1

B. Derivative of m[n] with respect to T2

The appendix gives iterative procedures for calculating magnetization derivatives with respect to T2 and M0. Each procedure uses the corresponding derivative relations and initial conditions.

B. Derivative of m[n] with respect to T2

C. Derivative of m[n] with respect to M0

Loading 1710.08062v8…