Elsevier

NeuroImage

Volume 183, December 2018, Pages 336-345
NeuroImage

MultiNet PyGRAPPA: Multiple neural networks for reconstructing variable density GRAPPA (a 1H FID MRSI study)

https://doi.org/10.1016/j.neuroimage.2018.08.032Get rights and content

Highlights

  • Novel acceleration/reconstruction method for non-lipid-suppressed 1H MRSI at 9.4T

  • Comparison of the introduced method to GRAPPA with retrospective lipid removal.

  • Highly accelerated metabolite maps from the human brain acquired at 9.4T

Abstract

Magnetic resonance spectroscopic imaging (MRSI) is a powerful tool for mapping metabolite levels across the brain, however, it generally suffers from long scan times. This severely hinders its application in clinical settings. Additionally, the presence of nuisance signals (e.g. the subcutaneous lipid signals close to the skull region in brain metabolite mapping) makes it challenging to apply conventional acceleration techniques to shorten the scan times. The goal of this work is, therefore, to increase the overall applicability of high resolution metabolite mapping using 1H MRSI by introducing a novel GRAPPA acceleration acquisition/reconstruction technique. An improved reconstruction method (MultiNet) is introduced that uses machine learning, specifically neural networks, to reconstruct accelerated data. The method is further modified to use more neural networks with nonlinear hidden layers and is then combined with a variable density undersampling scheme (MultiNet PyGRAPPA) to enable higher in-plane acceleration factors of R = 5.6 and R = 7 for a non-lipid suppressed ultra-short TR and TE 1H FID MRSI sequence. The proposed method is evaluated for high resolution metabolite mapping of the human brain at 9.4T. The results show that the proposed method is superior to conventional GRAPPA: there is no significant residual lipid aliasing artifact in the images when the proposed MultiNet method is used. Furthermore, the MultiNet PyGRAPPA acquisition/reconstruction method with R = 5.6 results in reproducible high resolution metabolite maps (with an in-plane matrix size of 64 × 64) that can be acquired in 2.8 min on 9.4T. In conclusion, using multiple neural networks to predict the missing points in GRAPPA reconstruction results in a more reliable data recovery while keeping the noise levels under control. Combining this high fidelity reconstruction with variable density undersampling (MultiNet PyGRAPPA) enables higher in-plane acceleration factors even for non-lipid suppressed 1H FID MRSI, without introducing any structured aliasing artifact in the image.

Introduction

Shortly after its introduction in 2002 (Griswold et al., 2002), the generalized partial parallel acquisition (GRAPPA) acceleration method was adopted by the magnetic resonance spectroscopy imaging (MRSI) community to accelerate the lengthy MRSI scans. Contrary to the sensitivity encoding (SENSE) (Pruessmann et al., 1999) acceleration method, GRAPPA does not require explicit knowledge of the receive coil sensitivity profiles. It is easy to implement and can be applied in any phase encoding direction. This makes it a suitable method for acceleration of MRSI sequences. GRAPPA was first incorporated in a localized MRSI sequence (Banerjee et al., 2006) in both phase-encoding directions to accelerate the scan by a net factor of R ∼ 2. Later on, GRAPPA (still with moderate acceleration factors) was used in combination with proton echo-planar spectroscopy imaging (PEPSI) sequences (Tsai et al., 2008; Sabati et al., 2014) to accelerate the MRSI studies even further.

Given that applying PEPSI sequences at ultra-high fields is challenging due to higher gradient strength and receive bandwidth requirements, one of the most promising MRSI techniques at ultra-high fields is the slice-selective non-lipid suppressed ultra-short TR and TE 1H FID MRSI sequence (Bogner et al., 2012; Boer et al., 2012; Nassirpour et al., 2018). Hence, the next natural step was to incorporate GRAPPA into this sequence to accelerate the scans even further. However, it was soon discovered that extra care should be taken in applying conventional parallel imaging (PI) methods such as GRAPPA to non-lipid suppressed 1H MRSI sequence. The reason for this is the presence of strong lipid signals near the skull (for brain metabolite mapping) which can be orders of magnitude stronger than the metabolites of interest inside the brain. Any remaining unresolved and structured aliasing artifact resulting from reconstruction errors in the conventional PI acceleration methods can completely distort the signal inside the brain and make quantification impossible. To overcome this, Hangel et al. (2015) suggested incorporating a double-inversion recovery (DIR) lipid suppression scheme into the 1H FID MRSI sequence. The DIR method is very effective for reducing the lipid signals and hence enabling higher PI acceleration factors, however, the major limitation is the strict specific absorption rate (SAR) limitations at ultra-high fields. As a result of this, the repetition time (TR) is prolonged by a significant factor which is only partially compensated by using higher GRAPPA acceleration factors. Alternatively, lipid contamination can be retrospectively removed. Bilgic et al. (2014) proposed using the spectra in the subcutaneous lipid region and decorrelating it with the spectra within the brain using L2 regularization. This was later applied to non-lipid-suppressed MRSI at 7T (Hangel et al., 2018).

The most recent development for accelerating ultra-high field MRSI using an ultra-short TR and TE 1H FID MRSI sequence was when Strasser et al. (2017) took another PI approach and used (2 + 1) D CAIPIRINHA to enable higher acceleration factors despite the presence of unsuppressed lipids. They accomplished this by controlling the aliasing pattern through optimizing the undersampling scheme. However, the high acceleration factor resulting from this method relies on simultaneous acquisition of signal from multiple slices, which limits the use of slice-wise B0 shim updating. It was also shown that the method sometimes still suffers from residual lipid contamination artifacts.

Given that the remaining lipid aliasing artifacts are a direct result of inaccuracies in the GRAPPA reconstruction process, and in the interest of enabling slice-wise dynamic B0 shim updating, in this study we take a different approach and try to enable high in-plane GRAPPA acceleration factors for non-lipid suppressed 1H FID MRSI by introducing a novel and more accurate GRAPPA acquisition and reconstruction scheme. The problem of B0 inhomogeneity is much more severe at ultra-high fields and thus it is important that good B0 shimming such as slice-wise dynamic B0 shim updating is possible. Furthermore, at higher field strengths, the B1+ inhomogeneity is more severe making it difficult to achieve uniform excitation and also higher B1 inhomogeneity makes quantification of metabolites more difficult. These problems that are present at ultra-high fields can introduce more artifacts than at lower field strengths.

Since the early years of GRAPPA, the imaging community has introduced several variations to the reconstruction process that would increase its accuracy. Among these advances was the realization that separating the reconstruction to the low and high frequency regions in k-space and having separate kernels for each, will increase the accuracy of the reconstruction and suppress the residual aliasing as well as noise amplification artifacts (Park et al., 2005; Miao et al., 2011). Park et al. (2005) further used this property to introduce 1D variable density sampling with higher acceleration factors in the outer k-space without losing accuracy. Another group (Wang et al., 2005) used a multi-line, multi-column interpolation approach for finding a more accurate GRAPPA reconstruction kernel. By increasing the number of training points, this method improves the quality of reconstruction. Huang et al. (2008) used an image-support based approach for more accurate data recovery.

Additionally, regularization has been proven to be advantageous in PI reconstruction techniques and hence (Qu et al., 2006) used Tikhonov and singular-value decomposition to regularize the GRAPPA reconstruction optimization problem to control the trade-off between noise and residual artifacts in the resulting image. In another approach, Huo and Wilson, 2008 introduced Robust GRAPPA, in which they assign weights to the training data in a way that would discount the contribution of the outliers to the coefficient estimation. Other groups used cross-validation (Nana et al., 2008) to better determine which neighboring lines and columns should be used in forming the reconstruction kernel. A more recent advance for improving the GRAPPA reconstruction was the introduction of nonlinear GRAPPA by Chang et al. (2012). All GRAPPA reconstruction methods so far assume a linear relationship between the acquired data points and the missing data, and hence, form a linear optimization problem to find the optimal reconstruction kernel. However, Chang et al. (2012) observed a nonlinear relationship between the acquired auto-calibration signal (ACS) and the missing data points in the presence of noise. Their reconstruction optimization problem takes these nonlinearities into account by introducing up to 2nd order polynomial terms into the system of equations used for kernel optimization. Their reconstruction method proved superior to conventional GRAPPA for higher acceleration factors.

The use of machine learning in MRI is not new. Yan and Mao, 1993 used neural networks to reduce truncation artifacts resulting from a truncated k-space. This filters the image domain using a trained nonlinear predictor. Neural networks were shown to, again, reduce truncation artifacts by Hui and Smith (1995) and for complex data (Smith and Hui, 1997a, Smith and Hui, 1997b). Neural networks are essentially a method for nonlinear regression. The input data is a matrix of N features and the features are mapped to a nonlinear space in a hidden layer. The output of the network is a linear combination of this hidden layer. The complexity of the NN can be increased by increasing the number of layers or the number of nodes in the layer. For reconstruction of undersampled data, neural networks have been used for parallel imaging in MRI (Sinha et al., 2007) which showed improvement over conventional GRAPPA and SENSE in imaging. Even more recently, a deep convolutional neural network by Schlemper et al. (2017) showed promising results for reconstructing highly accelerated images. However, to our knowledge, the application of machine learning for reconstructing spectroscopic data is limited.

In this work we present a novel and improved GRAPPA reconstruction method (MultiNet GRAPPA) that combines the advantages of regularization, cross-validation, and localized coil calibration, and accounts for nonlinearities by using multiple neural networks in the reconstruction process. We show the advantages of this approach over the conventional GRAPPA reconstruction for accelerating non-lipid suppressed ultra-short TR and TE 1H FID MRSI and further introduce a modified version combined with a variable density sampling scheme (MultiNet PyGRAPPA) that enables higher acceleration factors. Finally, we show the reproducibility of this approach for fast and high resolution metabolite mapping of the human brain at 9.4T.

Section snippets

Data acquisition

High-resolution MRSI data were acquired using a slice-selective 1H FID MRSI sequence (Henning et al., 2009; Bogner et al., 2012) with ultra-short TE and TR without any outer volume or lipid suppression schemes. An optimized 3-pulse water suppression scheme with a total duration of 62 ms was implemented in the sequence and optimized for a range of T1 values and B1 inhomogeneity levels between 50 and 150% (Nassirpour et al., 2018). Five healthy volunteers were scanned using a Siemens 9.4T

MultiNet GRAPPA

Fig. 4 shows the metabolite maps for NAA and Glutamate. The elliptically sampled data (R = 1), conventional GRAPPA reconstructed data (R = 2 × 2) with and without L2-regularization and MultiNet reconstructed data (R = 2 × 2) are shown in the figure for three subjects. Figure S2 of the supplementary materials shows the corresponding tCho maps. The advantage of the proposed MultiNet reconstruction over the conventional GRAPPA reconstruction can be seen from the residual aliasing artifacts present

Discussion

Metabolite mapping using 1H MRSI is a promising tool for assessing the spatial distribution of several metabolites in different regions of the brain in a non-invasive manner. To increase the clinical applicability of 1H MRSI, it is crucial to increase the SNR and in general the data quality as much as possible, but also decrease the lengthy scan times. The slice-selective ultra-short TE and TR non-lipid suppressed 1H FID MRSI sequence (Bogner et al., 2012; Boer et al., 2012; Nassirpour et al.,

Conclusion

This paper presents a novel acquisition and reconstruction scheme (MultiNet PyGRAPPA) for highly accelerated metabolite mapping using the non-lipid suppressed 1H FID MRSI sequence. The proposed method is an improvement on conventional GRAPPA, combining a variable density sampling scheme with multiple neural networks to predict the missing data points. The machine learning based reconstruction method proves superior to conventional GRAPPA and results in less noise and residual aliasing

Acknowledgements

This study was supported by the European Research Council Starting grant (project SYNAPLAST MR #679927) and the Horizon 2020 grant (project CDS_QUAMRI #634541).

References (38)

  • G. Hangel et al.

    Ultra-high resolution brain metabolite mapping at 7T by short-TR Hadamard-encoded FID-MRSI

    Neuroimage

    (2018)
  • Sahar Nassirpour et al.

    High and ultra-high resolution metabolite mapping of the human brain using 1H FID MRSI

    Neuroimage

    (2018)
  • N.I. Avdievich et al.

    Combined surface Loop/"Vertical" loop element improve receive performance of a human head transceiver array at 9.4T: an alternative to surface loop/dipole antenna combination

    NMR Biomed.

    (2018)
  • S. Banerjee et al.

    Fast magnetic resonance spectroscopic imaging at 3 Tesla using autocalibrating parallel technique

  • B. Bilgic et al.

    Fast image reconstruction with L2-regularization

    J. Magn. Reson. Imag.

    (2014)
  • M. Blaimer et al.

    2D-GRAPPA-operator for faster 3D parallel MRI

    Magn. Reson. Med.

    (2006)
  • V.O. Boer et al.

    Multi-slice 1H MRSI of the human brain at 7 Tesla using dynamic B0 and B1 shimming

    Magn. Reson. Med.

    (2012)
  • W. Bogner et al.

    High-resolution mapping of human brain metabolites by free induction decay 1H MRSI at 7 T

    NMR Biomed.

    (2012)
  • G.L. Chadzynski et al.

    Fast and efficient free induction decay MR spectroscopic imaging of the human brain at 9.4 Tesla

    Magn. Reson. Med.

    (2017)
  • Y. Chang et al.

    Nonlinear GRAPPA: a kernel approach to parallel MRI reconstruction

    Magn. Reson. Med.

    (2012)
  • P. Chang et al.

    Modeling real shim fields for very high degree (and order) B0 shimming of the human brain at 9.4 T

    Magn. Reson. Med.

    (2018)
  • M.A. Griswold et al.

    Generalized autocalibrating partially parallel acquisitions (GRAPPA)

    Magn. Reson. Med.

    (2002)
  • G. Hangel et al.

    Lipid suppression via double inversion recovery with symmetric frequency sweep for robust 2D-GRAPPA-accelerated MRSI of the brain at 7 T

    NMR Biomed.

    (2015)
  • A. Henning et al.

    Slice-selective FID acquisition, localized by outer volume suppression (FIDLOVS) for 1H-MRSI of the human brain at 7 T with minimal signal loss

    NMR Biomed.

    (2009)
  • F. Huang et al.

    High-pass GRAPPA: an image support reduction technique for improved partially parallel imaging

    Magn. Reson. Med.

    (2008)
  • Y. Hui et al.

    MRI reconstruction from truncated data using a complex domain backpropagation neural network

  • D. Huo et al.

    Robust GRAPPA reconstruction and its evaluation with the perceptual difference model

    J. Magn. Reson. Imag.

    (2008)
  • T. Kirchner et al.

    Mechanisms of SNR and line shape improvement by B0 correction in overdiscrete MRSI reconstruction

    Magn. Reson. Med.

    (2017)
  • D.C. Liu et al.

    On the limited memory BFGS method for large scale optimization methods

    Math. Program.

    (1989)
  • Cited by (18)

    • Quantitative T<inf>1</inf>-relaxation corrected metabolite mapping of 12 metabolites in the human brain at 9.4 T

      2022, NeuroImage
      Citation Excerpt :

      This work aims to provide a reference for quantitative data at UHF; with this study, other studies should be able to directly compare quantitative results acquired with accelerated techniques. This reference can provide valuable insight into how to utilize acceleration techniques such as highly-accelerated water referencing (Chang et al., 2018) or variable density under-sampling techniques (Nassirpour et al., 2018b, 2018a) for metabolite quantification, and therefore, make quantitative metabolic imaging with 1H MRSI more clinically feasible. MRSI quantification accuracy and precision at UHF could benefit from corrections for B1+ and B1− fields either by correcting for B1 fields while processing data or by implementing hardware solutions that correct B1 fields during data acquisition. .

    • Emerging methods and applications of ultra-high field MR spectroscopic imaging in the human brain

      2022, Analytical Biochemistry
      Citation Excerpt :

      However, MRSI also faces more challenges due to the large voxel size [77], or overwhelming lipid signals. In UHF MRSI, parallel imaging has been used extensively [73,78–82] and benefits especially from the more localized sensitivities of array coils [83]. Compressed sensing relies on the fact that the acquisition can be accelerated by random undersampling if data are compressible.

    • Achieving high-resolution <sup>1</sup>H-MRSI of the human brain with compressed-sensing and low-rank reconstruction at 7 Tesla

      2021, Journal of Magnetic Resonance
      Citation Excerpt :

      Hence, by combining LR and CS in our reconstruction we can take advantage of both properties to improve data quality and quantification. Spectral quality observable in Fig. 3 with the different spectra and assessed by the fitting residuals is unaffected by CS acceleration, in agreement with previous observations [65]. To the best of our knowledge, the first 3D images of metabolite distributions over extensive brain coverage were published by Duijn et al. [66] with low-resolution Cartesian encoding at 2T MRI scanner.

    • Imaging molecules

      2019, Journal of Magnetic Resonance
    View all citing articles on Scopus
    1

    These authors contributed equally to this work.

    View full text