Simultaneous QSM and Metabolic Imaging of the Brain Using SPICE: Further Improvements in Data Acquisition and Processing

Abstract

Purpose:

To achieve high-resolution mapping of brain tissue susceptibility in simultaneous QSM and metabolic imaging.

Methods:

Simultaneous QSM and metabolic imaging was first achieved using SPICE (SPectroscopic Imaging by exploiting spatiospectral CorrElation), but the QSM maps thus obtained were at relatively low-resolution (2.0×3.0×3.0 mm³). We overcome this limitation using an improved SPICE data acquisition method with the following novel features: a) sampling (k, t)-space in dual densities, b) sampling central k-space fully to achieve nominal spatial resolution of 3.0×3.0×3.0 mm³ for metabolic imaging, and c) sampling outer k-space sparsely to achieve spatial resolution of 1.0×1.0×1.9 mm³ for QSM. To keep the scan time short, we acquired spatiospectral encodings in EPSI trajectories in central k-space but in CAIPIRINHA trajectories in outer k-space using blipped phase encodings. For data processing and image reconstruction, a union-of-subspaces model was used, effectively incorporating sensitivity encoding, spatial priors and spectral priors of individual molecules.

Results:

In-vivo experiments were carried out to evaluate the feasibility and potential of the proposed method. In a 6-min scan, QSM maps of 1.0×1.0×1.9 mm³ resolution and metabolic maps at 3.0×3.0×3.0 mm³ nominal resolution were obtained simultaneously. Compared with the original method, the QSM maps obtained using the new method reveal fine-scale brain structures more clearly.

Conclusion:

We demonstrated the feasibility of achieving high-resolution QSM simultaneously with metabolic imaging using a modified SPICE acquisition method. The improved capability of SPICE may further enhance its practical utility in brain mapping.

INTRODUCTION

Quantitative susceptibility mapping (QSM) has been widely used for imaging tissue magnetic susceptibility in a range of applications,^1,2 such as detection of intracerebral hemorrhage,^3,4 measurement of brain oxygen consumption,^5–7 and assessment of neurodegenerative^8–11 and inflammatory diseases.^12,13 Complementary to QSM, MR spectroscopic imaging (MRSI) can map the spatial distribution of metabolites and neurotransmitters in the brain,^14–16 which is useful for studying brain metabolism and early detection and characterization of various diseases.^17–19

In the current practice, QSM and MRSI experiments are carried out separately using different acquisition sequences, which usually require a long scan time. Specifically, QSM data is often acquired using a multi-echo GRE sequence, which takes around 3–5 minutes to achieve a resolution on the order of 1 mm;^1,20 MRSI data is usually acquired using a CSI sequence, which takes more than 20 minutes to cover one slice with resolution around 1 cm.^16,21 Fast simultaneous QSM and MRSI was demonstrated recently using SPICE (SPectroscopic Imaging by exploiting spatiospectral CorrElation).^22–25 However, the spatial resolution of the QSM maps obtained using the basic SPICE sequence was limited to around 2.0 × 3.0 × 3.0 mm³. While this resolution is impressive for metabolic imaging, it is significantly lower than what is provided by conventional QSM techniques.²² Low-resolution QSM may not be able to capture microstructure features in the brain and also result in a loss of susceptibility contrast and errors in susceptibility estimation.²⁶ One approach to address this limitation is to increase the number of spatiospectral encodings, but at the expense of lengthening the scan time. For example, increasing the resolution from 2.0 × 3.0 × 3.0 mm³ to 1.0 × 1.0 × 1.9 mm³ (for both QSM and metabolic imaging) would increase the acquisition time from 5 minutes to 45 minutes, which is not acceptable in practice.

In this work, we propose an improved method to obtain high resolution QSM in 1.0 × 1.0 × 1.9 mm³ simultaneously with metabolic imaging in 3.0 × 3.0 × 3.0 mm³ in a 6-min scan with an FOV of 240 × 240 × 72 mm³. The proposed method samples (k, t)-space in dual densities; more specifically, the central k-space region is fully sampled to achieve 3.0 × 3.0 × 3.0 mm³ nominal spatial resolution for metabolic imaging while outer k-space is sampled sparsely to achieve 1.0 × 1.0 × 1.9 mm³ resolution for QSM. To keep the scan time short, the proposed method acquires spatiospectral encodings using EPSI trajectories in the central k-space region but uses CAIPIRINHA (Controlled Aliasing in Parallel Imaging Results in Higher Acceleration) trajectories in the outer k-space region using blipped phase encodings.^27–29 For data processing and image reconstruction, a union-of-subspaces model is used, effectively incorporating sensitivity encoding, spatial priors and spectral priors of individual molecules. In vivo simultaneous QSM/MRSI experiments have been carried out on both healthy subjects and tumor patients to demonstrate the improved performance achieved by the proposed method.

METHODS

Data Acquisition:

The proposed data acquisition sequence keeps the essential features of the basic SPICE sequence, as shown in Fig. 1(a),^22–25 to enable fast simultaneous acquisition of QSM and MRSI data: (1) elimination of the water and lipid suppression pulses, and (2) acquisition of FID signals in EPSI-based trajectories with ultrashort TE (1.6 ms) and short TR (160 ms) for rapid acquisition of spatiotemporal encodings. Figure 1(b) illustrates the (k, t)-space sampling pattern of the proposed data acquisition method. Note that the central k-space region is sampled as was done in the original SPICE technique,²² which provides sufficient coverage to achieve 2.0 × 3.0 × 3.0 mm³ resolution for QSM (and 3.0 × 3.0 × 3.0 mm³ nominal resolution for metabolic maps). The proposed method extends the basic SPICE sequence by collecting another set of spatiospectral encodings in outer k-space in the same spin steady state. This set of data is acquired rapidly with a scan time on the order of 1 minute. This efficiency is achieved with highly sparse sampling, exploiting the fact that the outer k-space data are used for QSM mapping only. More specifically, the outer k-space region is uniformly undersampled along ky, while 2D spatiotemporal CAIPIRINHA trajectories are used to cover the (kz, t)-space. In sampling (kz, t)-space, both the odd and even echoes of the gradient echo trains are utilized to increase data acquisition efficiency.

Figure 1:

(a) Sequence diagram of the proposed acquisition method: FID signals are acquired in EPSI trajectories without water and lipid suppression (TE = 1.6 ms, TR = 160 ms, echospace = 1.76 ms); blip phase gradients (shown in red) are used to cover the outside k-space region. (b) (k, t)-space sampling pattern: The central region is fully sampled while the outside k-space region is sparsely sampled using 2D CAIPIRINHA trajectories.

In our current implementation of the proposed method, for a typical 3D scan (FOV = 240 × 240 × 72 mm³; TR = 160 ms) of 6 minutes, the data acquisition scheme acquires 126 × 78 × 24 spatial encodings in the central k-space region in 5 minutes and 126 × 216 × 72 encodings in outer k-space in 1 minute. In the outer k-space region, the ky-dimension is undersampled by a factor of 3 while the (kz, t)-space is covered using CAIPIRINHA trajectories with a factor of 12 undersampling. Therefore, as compared with the central k-space region, the outer k-space region is sampled with an acceleration factor of 36, which is permissible since only the water component is used for QSM. The effective k-space sampling size is 216 × 72 × 126, providing a resolution of 1.0 × 1.0 × 1.9 mm³ for the QSM component. So, the proposed data acquisition scheme enhances the spatial resolution of the QSM component by a factor of 9 with an extra 1-min scan time.

To demonstrate the feasibility of the proposed method, we implemented the pulse sequence on 3T Siemens Prisma/Skyra scanners (Siemens Medical Solutions, Erlangen, Germany). In vivo experiments on three healthy subjects were performed with approval by the Institutional Review Board of the University of Illinois. SPICE scans on four tumor patients were approved by the Institutional Review Board of the Fifth People’s Hospital of Shanghai; written informed consents were obtained from all the participants. The scan protocol included a traditional MPRAGE scan (FOV = 240 × 240 × 192 mm³, matrix size = 256 × 256 × 192, TE = 2.29 ms, TI = 900 ms, TR = 1900 ms) and a SPICE scan using the proposed method with the parameters described above (other parameters include: flip angle = 27˚, echo space = 1.76 ms, readout bandwidth = 167 kHz). For the tumor patients, gadolinium diethylenetriaminepentaacetic acid (Gd-DTPA) was intravenously administered for contrast enhancement of the MPRAGE images as part of the clinical protocol for these patients.

Image Reconstruction:

Given the new data acquisition scheme, the key data processing task lies in reconstructing the water signals in high resolution utilizing the extended k-space data. This was accomplished by using the union-of-subspaces model as in the original SPICE method. More specifically, we represented the spatiotemporal signals of water, lipids, and metabolites, each in a very low-dimensional subspace^{22,25,30–33}:

$${\rho }_{\text{MRSI}}(\mathit{\text{r}},t)={\sum }_{l_w=1}^{L_w}{U}_{l_w}(\mathit{\text{r}})V_{l_w}(t)+{\sum }_{l_f=1}^{L_f}{U}_{l_f}(\mathit{\text{r}})V_{l_f}(t)+{\sum }_{l_m=1}^{L_m}{U}_{l_m}(\mathit{\text{r}})V_{l_m}(t),$$ [1]

where $V_{l_w},\mathrm{ }{V}_{l_f},\mathrm{ }{V}_{l_m}$ are temporal basis functions for the water, lipid and metabolite signals, respectively, and $U_{l_w},\mathrm{ }{U}_{l_f},\mathrm{ }{U}_{l_m}$ are their corresponding spatial coefficients.

With this model, reconstruction of the high-resolution water and lipid signals was accomplished by solving the following optimization problem (ignoring the metabolite components):

$$\underset{\left\{U_{l_w},U_{l_f}\right\}}{\text{min}}d-\Omega \mathcal{F}BS{\left(M_w\odot {\sum }_{l_w=1}^{L_w}{U}_{l_w}{V}_{l_w}+M_f\odot {\sum }_{l_f=1}^{L_f}{U}_{l_f}{V}_{l_f}\right)}_2^2+R\left(U_{l_w},U_{l_f}\right)$$ [2]

where d denotes the measured data from both the central and outer k-space regions, $\mathrm{\Omega },\mathcal{F},B,S$ are operators that represent sampling in (k, t)-space, Fourier transformation, B₀ field inhomogeneity effect and sensitivity encoding, respectively. M_w and M_f represent the spatial supports for the water and lipid signals, and R is an edge-preserving regularization operator.^24,34 The temporal basis functions of the water and lipid signals, $\left\{V_{l_w}\right\}$ and $\left\{V_{l_f}\right\}$, were pre-determined from the central fully sampled k-space using a singular value decomposition (SVD)-based method with field correction.^23,32 The spatial coefficients estimated from equation [2] were then used to synthesize the high-resolution water and lipid signals as follows

$${\rho }_w={\sum }_{l_w=1}^{L_w}{U}_{l_w}{V}_{l_w},{\rho }_f={\sum }_{l_f=1}^{L_f}{U}_{l_f}{V}_{l_f}$$ [3]

In this reconstruction process, the coil sensitivity maps and the field inhomogeneity map were first estimated from the central low-resolution k-space data, which contained truncation effects that may affect the resulting reconstruction. Besides, the signal distortion caused by the partial volume effect, especially in the frontal region where large field inhomogeneity exists due to susceptibility changes, can affect the estimation of the subspace structures. To overcome these problems, we updated the estimation of the coil sensitivity maps, the field inhomogeneity map and the spectral basis functions by utilizing both the extended k-space data and the initial reconstruction. We also removed the lipid signals obtained from the initial reconstruction of the measured data, which led to a simpler reconstruction model. We then reconstructed the water signals again with a higher model order to capture subtle image features. After the water signals were reconstructed in high-resolution, QSM maps were calculated using an existing processing pipeline which includes field estimation using HSVD,³⁵ background field removal by solving the Laplacian boundary value problem,³⁶ and solving the dipole-inversion model incorporating anatomical spatial priors.^22,37 The spatiospectral reconstruction of the metabolites were done using the original SPICE processing pipeline.^{23–25,30–33}

RESULTS

To investigate the effect of sampling trajectories on reconstruction quality, one full high-resolution dataset was collected using the proposed sequence without sparse sampling of k-space. Three practical trajectories were designed and tested on the experimental data by retrospective undersampling. The three sampling trajectories had the same acceleration factors both spatially and temporally and the reconstruction errors were calculated using the reconstruction from the full dataset as the ground truth. A set of representative reconstruction results are shown in Fig. 2. As can be seen, the reconstructions using the proposed method were reasonably robust against different sampling trajectories. The third sampling trajectory, which was implemented in our current sequence, yielded the smallest reconstruction error.

Figure 2:

Comparison of reconstruction results from three different (kz, t) sampling trajectories. As can be seen, the proposed method is relatively insensitive to the choice of sampling trajectories although reconstruction quality is trajectory-dependent.

To evaluate the high-resolution QSM capability of the proposed method, we used the standard GRE-based high-resolution QSM methods as our gold standard (reference).¹ Our goal was to test if we could produce the standard high-resolution QSM results in the MRSI setting. To this end, we collected both GRE-QSM and SPICE data from a healthy human subject. The GRE-QSM data were collected using a standard multi-echo GRE sequence with similar parameters as the SPICE sequence (FOV = 240 × 240 × 72 mm³, resolution = 1.0 × 1.0 × 1.9 mm³, echo number = 20, echo space = 1.76 ms, TE = 1.60 ms, TR = 60 ms). Figure 3 shows a set of representative results comparing QSM maps obtained from the standard QSM method, our proposed method, and the previous low-resolution SPICE-QSM.²² As can be seen from the zoom-in regions in Fig. 3, some vein structures were lost in the low-resolution QSM maps due to the partial volume effect, which were recovered nicely in the high-resolution QSM maps produced by the proposed method. These high-resolution structures revealed by the proposed method were also consistent with those in the QSM map from the standard QSM method.

Figure 3:

Comparison of QSM maps obtained from a healthy subject using (a) low-resolution SPICE-QSM, (b) the proposed method, and (c) standard high-resolution QSM method, respectively. Note the proposed method and the standard high-resolution QSM method produced comparable QSM maps, capturing the microstructures (e.g. veins) in the brain well. The improvement over the existing low-resolution SPICE-QSM method is also noticeable.

In addition to resolution improvement for QSM, the high-resolution water signals acquired also provide a high-resolution field map, which helped minimize the field inhomogeneity effects on reconstruction of the metabolite signals, as shown in the Supporting Information. Figure 4 shows a set of representative results from a healthy subject, which include the high-resolution (1.0 × 1.0 × 1.9 mm³) QSM maps and the concentration maps of several metabolites (including NAA, Cr and Cho, reconstructed in 3.0 × 3.0 × 3.0 mm³ resolution), which were obtained from the 6-min scan. The QSM maps and the metabolite maps from several selected slices across the brain are shown on the left and the spatially resolved spectra are displayed on the right, illustrating the high spatiospectral quality of the reconstruction results.

Figure 4:

A set of representative results obtained from a healthy subject using the proposed method in a 6-min scan. The high-resolution QSM maps with a spatial resolution of 1.0 × 1.0 × 1.9 mm³ were reconstructed using all the k-space data while the metabolic maps were reconstructed in a nominal resolution of 3.0 × 3.0 × 3.0 mm³ using only the central-space data.

Another set of results from a tumor patient are displayed in Fig. 5. As can be seen from the contrast enhanced T₁-weighted images, there are two metastatic lesions in the brain. From the metabolic maps and the localized spectra, one can clearly see a reduction in NAA and an elevation in choline in both lesions, which is consistent with the literature.^38,39 In the QSM maps, high tissue susceptibility can be observed in one of the lesions but not the other, which may indicate the differences of blood deposition in these two lesions.⁴⁰

Figure 5:

Experimental results from a tumor patient using the proposed method in a 6-min scan. The experiment also collected a set of contrast-enhanced T₁-weighted images. Localized spectra in the tumor (marked by blue box in T₁-weighted images) and its contralateral region (marked by red box in T₁-weighted images) clearly show the reduction in NAA and an in increase in Cho in the tumor region.

DISCUSSION

This paper presents a major extension of SPICE for fast simultaneous QSM and metabolic imaging, reaching a better trade-off in addressing the data acquisition requirements for metabolic imaging and QSM. Keeping the scan time short (e.g., around 6 minutes), the resolution of QSM is improved by a factor of 9 without affecting the quality of the MRSI signals.

The acceleration factor in the outer k-space region is 36; and we have demonstrated good reconstruction performance using our proposed method leveraging a union-of-subspaces model to incorporate both spatial and spectral priors. Keeping the scan time the same, we expect that the reconstruction results can be further improved by a) using a more optimal sparse sampling scheme, and b) using stronger prior information.

One way to improve our current sampling scheme is to take better advantage of the spatiotemporal correlation of (k, t)-space signals. Currently, we sample (kz, t)-space using 2D CAIPIRINHA trajectories; given the flexibility we have in designing the sampling trajectories using blipped gradients, we could simultaneously sample (ky, kz, t)-space using 3D CAIPIRINHA trajectories, which may further improve our current reconstruction results.²⁹

The spatial prior information used in the proposed method is a weak weighting function extracted from a single structural image (e.g. MPRAGE image); we could use deep learning-based methods to generate stronger priors.^41–43 Such methods would become possible after more SPICE data become available and can be used as training data.

CONCLUSIONS

We proposed a new method to achieve high-resolution QSM for fast simultaneous QSM and MRSI. Experimental results demonstrated that the proposed method can achieve simultaneous QSM at 1.0 × 1.0 × 1.9 mm³ nominal resolution and metabolite maps at 3.0 × 3.0 × 3.0 mm³ nominal resolution in a single 6-min scan. The proposed method would enhance the practical usefulness of SPICE in mapping brain metabolism and tissue magnetic properties.

Supplementary Material

Supp info

Figure S1. Singular value decay curves of the Casorati matrix formed with SPICE data before (green) and after (blue) field inhomogeneity correction. The faster decay due to field inhomogeneity correction implies improved partial separability of the SPICE data.

Figure S2. NAA maps from SPICE reconstructions using field maps at different resolutions: 1) no field correction, 2) 10×10×10 mm³ (typical CSI resolution), c) 2.0×3.0×3.0 mm³ (resolution of metabolite signals), and d) 1.0×1.0×1.9 mm³ (resolution of water signals; the proposed method). Note the signal cancelation effect in the frontal regions when low-resolution field correction was used.

Figure S3. Simulation study of the susceptibility effect in a tumor. A field map was generated with large susceptibility in the tumor (around 40 Hz). Reconstructions were performed with: (a) field correction using a high-resolution field map (1.0×1.0×1.9 mm³; resolution of the water image acquired using the proposed method), (b) no field correction, (c) field correction using a low-resolution field map (2.0×3.0×3.0 mm³, resolution of the metabolite signals).