Multiscale reconstruction for MR fingerprinting

Abstract

Purpose: To reduce the acquisition time needed to obtain reliable parametric maps with Magnetic Resonance Fingerprinting.

Methods: An iterative-denoising algorithm is initialized by reconstructing the MRF image series at low image resolution. For subsequent iterations, the method enforces pixel-wise fidelity to the best-matching dictionary template then enforces fidelity to the acquired data at slightly higher spatial resolution. After convergence, parametric maps with desirable spatial resolution are obtained through template matching of the final image series. The proposed method was evaluated on phantom and in vivo data using the highly undersampled, variable-density spiral trajectory and compared with the original MRF method. The benefits of additional sparsity constraints were also evaluated. When available, gold standard parameter maps were used to quantify the performance of each method.

Results: The proposed approach allowed convergence to accurate parametric maps with as few as 300 time points of acquisition, as compared to 1000 in the original MRF work. Simultaneous quantification of T1, T2, proton density (PD), and B0 field variations in the brain was achieved in vivo for a 256 × 256 matrix for a total acquisition time of 10.2 s, representing a three-fold reduction in acquisition time.

Conclusion: The proposed iterative multiscale reconstruction reliably increases MRF acquisition speed and accuracy. Magn Reson Med 75:2481-2492, 2016. © 2015 Wiley Periodicals, Inc.

Keywords: Compressed Sensing; Fingerprinting; Multiscale Image Reconstruction; Parameter Mapping.

Figure 1

Illustration of the iterative multi-scale algorithm: Starting from the top, the zero-filled measured k-space is Gaussian weighted at initialization. Step 1: y⁽ⁱ⁻¹⁾ is inverse-Fourier transformed, yielding x⁽ⁱ⁾. Step 2: MRF template matching is performed to denoise image series. If x⁽ⁱ⁾ had converged, stop here and output corresponding maps. Step 3: The denoised image series is Fourier transformed, yielding y⁽ⁱ⁾. Step 4: A new iteration starts, with a weaker Gaussian weighting of the original data and its direct substitution for the acquired elements of y⁽ⁱ⁾ at their sampled locations after Cartesian gridding.

Figure 2

MRF sequence parameters. a. Sequence of pseudo random TR values. b. Sequence of FA values for each excitation pulse. c. example VDS trajectory. Only measurements sampled along the full line were used for image reconstruction.

Figure 3

a. Comparison of T1 and T2 maps (upper and lower row, respectively) obtained after convergence of IMS-MRF (3^rd from left) purely iterative MRF (2^nd from left) to their respective ground truth (left). Error maps shown on the right. b. T2 normalized RMSE evolution with each iteration of IMS-MRF (light blue) and purely iterative MRF (black). c. From left to right, T2 maps obtained after initialization, iteration 3, iteration 6 and the last iteration of IMS-MRF (top row) and purely iterative MRF.

Figure 4

Normalized RMSE evolution with total acquisition time for each parameters. Original MRF values are displayed in red squares, IMS-MRF in blue circles and IMS-MRF with WV denoising in green triangles.

Figure 5

Maps of T1, T2, ΔB₀ and PD obtained with L=500 (5.1 s total acquisition) for Original MRF (top row), IMS-MRF (middle row) and IMS-MRF with WV denoising (bottom row). Reductions of residual errors in T2 and ΔB₀ maps for IMS-MRF with WV are indicated by arrows.

Figure 6

T2 maps (top row) and corresponding error maps (bottom row) obtained with original MRF (a) and IMS-MRF (b) using the first 300 TRs (left), 500 TRs (middle) and 1000 TRs (right).

Figure 7

Measured T1 (left) and T2 (right) comparison between SE and MRF measurements for a. 1000 TRs (10.2 s total acquisition time) b. 300 TRs (3.0 s total acquisition time) c. Evolution of the concordance coefficient with length of acquisition.

Figure 8

In vivo images from the 55th TR (left column), 335^h TR (middle column) and 717^th TR (right column) used to perform template matching with the original MRF method (top row) and IMS-MRF (bottom row) with 500 TRs. The duration of each TR is shown in parenthesis.

Figure 9

In vivo parameter maps from the original MRF (top row) and IMS-MRF methods (bottom row) for L =500 (a), 1000 (b) and 3000(c). The NRMSE of the L=500 and L=1000 maps compared to their L=3000 counterparts are shown as inset.