Ultrahigh-Resolution Metabolic Imaging at 9.4 Tesla
Fan Lam1, Hanbing Lu2, Yihong Yang2, Bryan Clifford1,3, Chao Ma1, Gene E Robinson4, and Zhi-Pei Liang1,3

1Beckman Institute, University of Illinois at Urbana-Champaign, Urbana, IL, United States, 2Neuroimaging Research Branch, National Institute on Drug Abuse, Baltimore, MD, United States, 3Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States, 4Carl R. Woese Institute for Genomic Biology, University of Illinois at Urbana-Champaign, Urbana, IL, United States


We present a multislice short-TE 1H-MRSI method to achieve fast, ultrahigh-resolution metabolic imaging of rats on a 9.4 Tesla animal scanner. The proposed method uses a subspace-based hybrid data acquisition strategy and a low-rank-model-based image reconstruction scheme. In vivo experiments have been performed to demonstrate the feasibility of the proposed method. We are able to produce high-SNR, spatially resolved metabolic profiles from the rat brain with 1x1x2mm3 nominal resolution in 16 minutes.


Metabolic imaging of small animals using MRSI is an important tool for studying basic neurobiological processes and various disease models (e.g., cancer, neurological disorders and psychiatric diseases), and for developing new therapeutic procedures and their preclinical translation.1 The development of advanced ultrahigh field animal scanners presents new opportunities in tackling the sensitivity and resolution challenges associated with animal imaging. However, due to the restricted tradeoffs of SNR, imaging speed, and resolution for MRSI, most metabolic studies still limit to single-voxel spectroscopy or slow phase-encoding-based chemical-shift imaging schemes.2,3 We present here a new method to achieve fast, ultrahigh-resolution 1H-MRSI on a 9.4T animal system, using subspace-based data acquisition and image reconstruction strategies. The proposed method is able to produce high-SNR, spatially resolved metabolic profiles from the rat brain with 1x1x2mm3 nominal resolution in 16 minutes.


Data Acquisition

To achieve accelerated acquisition for high-resolution MRSI, we use a special hybrid sampling strategy. Specifically, we developed a dual-density, dual-speed echo-planar spectroscopic imaging (EPSI) sequence on a 9.4T Bruker system (ParaVision 6.0.1). The sequence contains a low-bandwidth (BW), low-resolution EPSI scan to generate a set of high-SNR and spectrally fully-sampled training data, and a high-BW, high-resolution EPSI scan to generate a set of imaging data. This strategy is based on a low-dimensional subspace model that exploits the spatiospectral correlation in the spatiospectral function of interest.4 More specifically, the training data is for determining the subspace structure while the imaging data is for final spatiospectral reconstruction (See Image Reconstruction).

A multislice encoding scheme was used to achieve large brain coverage and avoid the truncation effects for 3D encoding due to limited encodings along the slice direction. Short-TE spin-echo acquisition was used to further improve the SNR. In addition, since most spectral components of interest for 1H-MRSI reside in 0-4ppm, the center frequency was adjusted to 2.3ppm, which enables a reduction of the sampling bandwidth due to reduced spectral bandwidth requirement.

Image Reconstruction

The nuisance water and lipid signals were first removed from both the training and imaging data, using the method in 5. The nuisance-signal-removed data, denoted as $$$s(\mathbf{k},t)$$$, was then modeled as

\begin{eqnarray*} s(\mathbf{k},t) & = & \int\left(\sum_{l_{m}=1}^{L_{m}}u_{l_{m}}(\mathbf{r})v_{l_{m}}(t)+\sum_{l_{b}=1}^{L_{b}}u_{l_{b}}(\mathbf{r})v_{l_{b}}(t)\right)e^{-i2\pi\Delta f(\mathbf{r})t}e^{-i2\pi\mathbf{k}\mathbf{r}}d\mathbf{r} + \eta(\mathbf{k},t) \\ \end{eqnarray*}

where $$$\sum_{l_m=1}^{L_m}u_{l_m}(\mathbf{r})v_{l_m}(t)$$$ and $$$\sum_{l_{b}=1}^{L_{b}}u_{l_{b}}(\mathbf{r})v_{l_{b}}(t)$$$ are subspace models for the metabolite and baseline signals, respectively, $$$L_m$$$ and $$$L_b$$$ the corresponding model orders (typically a small number), $$$\Delta f(\mathbf{r})$$$ the B0 map and $$$\eta(\mathbf{k},t)$$$ the noise. The baseline component was incorporated for very short-echo acquisitions. This model significantly reduces the degrees-of-freedom, making high-SNR and high-resolution reconstruction possible. Specifically, $$$\left\{v_{l_m}(.)\right\}$$$ and $$$\left\{v_{l_b}(.)\right\}$$$ were estimated from the training data4,6. Image reconstruction was then done by estimating $$$\left\{u_{l_m}(.)\right\}$$$ and $$$\left\{u_{l_{b}}(.)\right\}$$$ using the following regularized least-squares formulation (with proper discretization)

\begin{eqnarray}\hat{\mathbf{U}}_{m},\hat{\mathbf{U}}_{b} & = & \arg\underset{\mathbf{U}_{m},\mathbf{U}_{b}}\min\left\Vert \mathbf{s}-\mathcal{F}_{\Omega}\left\{ \mathbf{B}\odot\left(\mathbf{U}_{m}\mathbf{V}_{m}+\mathbf{U}_{b}\mathbf{V}_{b}\right)\right\} \right\Vert _{2}^{2}\nonumber \\ & & +\lambda_{1}\left\Vert \mathbf{D}_{w}\mathbf{U}_{m}\right\Vert _{F}^{2}+\lambda_{2}\left\Vert \mathbf{U}_{b}\mathbf{V}_b\mathbf{F}\mathbf{D}_{f}\right\Vert _{F}^{2},\label{eq:recon}\\\nonumber \end{eqnarray}

where the first term measures data fidelity, $$$\mathbf{B}$$$ models the B0 field inhomogeneity, $$$\mathcal{F}_{\Omega}$$$ is a Fourier encoding operator and $$$\mathbf{s}$$$ contains the imaging data. The regularization terms enforce spatial edge-preserving smoothness on the metabolite signal and spectral smoothness on the baseline signal, respectively. $$$\mathbf{D}_w$$$ and $$$\mathbf{D}_f$$$ perform spatial and spectral finite differences.


Data were acquired from a healthy rat brain (using a receive-only surface coil). The imaging parameters were: FOV = 40x40mm2, slice thickness = 2mm, TR/TE = 2200/15ms, the training data matrix size = 16x16, the imaging data matrix size = 40x40 (1x1mm2 nominal resolution), and water suppression BW = 180Hz. The readout BWs for the training and imaging data were 100 and 200kHz, respectively. The acquisition time for the imaging data was about 16min. The B0 map was from an EPSI scan without water suppression. The experimental setup is illustrated in Fig. 1. Four slices were acquired. $$$L_m = 8$$$ and $$$L_b=4$$$ were used for reconstruction based on SVD analysis. Figure 2 shows representative spectra from three voxels. As can be seen, the proposed method is able to produce high-SNR spectra, resolving regional differences of the metabolites. The NAA and glutamate maps (Fig. 3) also have very high quality (e.g., identifying features such as ventricle where low signal is expected). A comparison to Fourier reconstruction of the original EPSI data further demonstrates the capability of the proposed method (Fig. 4).


We have developed an ultrahigh-resolution 1H-MRSI method on a 9.4T animal scanner using subspace-driven acquisition and reconstruction. In vivo results demonstrate the feasibility of the proposed method in achieving fast, high-SNR and high-resolution metabolic imaging of the rat brain. With further validation and integration with spectral quantification, the proposed method should provide a useful tool for metabolic studies on various animal models.


This work was supported in part by NIH-R21EB021013-01 and the Beckman Institute Postdoctoral Fellowship.


1. Koo V, Hamilton PW, Williamson K. Non-invasive in vivo imaging in small animal research. Cell Oncol. 2006;28:127-39.

2. Tkac I, Henry PG, Andersen P, Keene CD, Low WC, Gruetter R. Highly Resolved In Vivo 1H NMR Spectroscopy of the Mouse Brain at 9.4 T. Magn Reson Med. 2015;52:478-484.

3. Mlynarik V, Kohler I, Gambarota G, Vaslin A, Clarke PG, Gruetter R. Quantitative proton spectroscopic imaging of the neurochemical profile in rat brain with microliter resolution at ultra-short echo times. Magn Reson Med. 2008;59:52-58.

4. Lam F, Ma C, Clifford B, Johnson CL, Liang ZP. High-resolution 1H-MRSI of the brain using SPICE: Data acquisition and image reconstruction. Magn Reson Med. 2015; In Press.

5. Ma C, Lam F, Johnson CL, Liang ZP. Removal of nuisance signals from limited and sparse 1H MRSI data using a union-of-subspaces model. Magn Reson Med. 2015; In Press.

6. Ma C, Lam F, Ning Q, Johnson CL, Liang ZP. High-resolution 1H-MRSI of the brain using short-TE SPICE. ISMRM 2015.


Figure 1: Experimental setup of the multislice EPSI acquisition. Four different slices were acquired covering a large portion of the brain. Six outer volume suppression bands were included for subcutaneous fat suppression.

Figure 2: Spectra from voxels located in three different brain regions. All spectra were normalized with respect to the highest peak in the first spectrum. As can be seen, the spectra from the cerebral cortex and the hippocampus show significantly higher levels of NAA and glutamate, consistent with existing literature.

Figure 3: Reconstructed NAA and glutamate (Glu) maps for one slice produced by the proposed method. Each map was normalized individually. The image on the most left shows the T2-weighted anatomical image (obtained by a RARE sequence) for the same slice.

Figure 4: A comparison of reconstructed spectra from the original EPSI data (middle column, Fourier reconstruction) and the proposed method (right column). The image on the left shows the anatomical image with the location of the voxel. As can be seen, significant SNR improvement is achieved by the proposed method.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)