Synopsis

Keywords: Spectroscopy, Quantitative Imaging

J-resolved MRSI is a powerful molecular imaging tool for measuring brain metabolites, neurotransmitters and other important biophysical parameters. The inherent SNR challenge of MRSI and prolonged scan time for multi-TE data limit the imaging resolution. This work presents a brand-new capability of whole-brain multiparametric, quantitative MRSI, by integrating a fast-scanning J-resolved MRSI sequence with SNR-efficient multi-band excitation, task-specific experiment designs, subspace imaging and optimized parameter estimation. Experimental studies and initial validation were performed to demonstrate this capability for high-resolution metabolite, neurotransmitter and metabolite T2 mapping from a single scan.

Introduction

J-resolved MRSI is a potentially powerful molecular imaging tool for improved detection and mapping of brain metabolites, neurotransmitters and other important biophysical parameters (e.g., T2), offering a rich set of potential biomarkers for disease research^1-3. However, existing J-resolved MRSI techniques are limited by low resolutions, limited SNR, small brain coverage and/or sub-optimal experimental design. Model-based imaging^4-6, e.g., sparse and low-rank models, have shown great potential in addressing these challenges. Moreover, optimized experimental designs, e.g., nonuniform TE sampling, have demonstrated improved metabolite quantification and parameter estimation^7-9. In this work, we present a brand-new capability of whole-brain multiparametric, quantitative MRSI, by integrating a fast-scanning J-resolved MRSI sequence with SNR-efficient multi-band excitation, task-specific experiment designs, subspace imaging and optimized parameter estimation. Experimental studies were performed to demonstrate this capability for high-resolution metabolite, neurotransmitter and metabolite T2 mapping from a single scan. T2 estimates from the proposed method were validated against standard single-voxel spectroscopy, which shows excellent consistency.

Theory and Methods

Proposed accelerated J-resolved MRSI acquisition
Recently, SPICE-based rapid acquisition strategies have demonstrated the capability of fast, high-resolution J-resolved MRSI^8-10. Multi-band excitation has also been implemented for single-TE, spectral editing MRSI¹¹ to achieve larger brain coverage. Inspired by these technical advancements, we proposed a fast sequence integrating multi-slab excitation, rapid spatiospectral encoding with (k,t,TE)-space sparse sampling and a new multi-slab interleaved water imaging acquisition. More specifically, two 3D slabs were sequentially excited and encoded in one TR. A pair of slab-selective adiabatic refocusing pulses were used for each slab to minimize CSDEs. This excitation was repeated for different TEs with TE-dependent (k_y,k_z)-undersampling for further acceleration at high resolutions. After MRSI encoding for each slab, field-drift navigators and water spectroscopic imaging data were interleaved following a small flip-angle (e.g., 10^o) water excitation. A blipped phase encoding strategy was used for the water imaging data which resulted in a more extended k-space coverage than the MRSI data. These water data can be used for tracking and correcting field drift, B₀ mapping and coil sensitivity estimation for interpolating the sparse MRSI data. The proposed sequence is illustrated in Fig. 1.
Building our own and other’s investigations on optimizing experimental design for J-resolved MRSI^7,8,12, different optimal combinations of TEs can be chosen for specific tasks, e.g., 2-TE combination with [65,80]ms for separating GABA and Glx signals¹² and 4-TE combination with [35,200,245,275]ms for estimating T2s of NAA, creatine (Cr) and choline (Cho)¹³. Balancing both tasks, a 4-TE combination of [35,65,80,245] ms can be chosen for simultaneous metabolite, neurotransmitter and T2 mapping.

Spatiospectral reconstruction using learned subspace with data-driven adaptation
Two important data processing challenges emerge, one on reconstructing high-SNR multi-TE spatiospectral functions from the rapidly collected noisy data and the other on accurately and reliably extracting quantitative parameters from the reconstruction. For reconstruction, we model the MRSI signals ($$$\mathbf{\rho}\left(\mathbf{r},t_{2},t_{1}\right)$$$) using a multi-TE union-of-subspaces (UoSS) model^9,14:
$$ \begin{aligned}\mathbf{\rho}\left(\mathbf{r},t_{2},t_{1}\right)=\sum_{l_{wat}=1}^{L_{wat}}c_{l_{wat}}(\mathbf{r})v_{l_{wat}}\left(t_{2},t_{1}\right)+\sum_{l_{lip}=1}^{L_{lip}}c_{l_{lip}}(\mathbf{r})v_{l_{lip}}\left(t_{2},t_{1}\right)+\sum_{l_{\text {met}}=1}^{L_{\text {met}}}c_{l_{\text {met}}}(\mathbf{r}) v_{l_{\text {met}}}\left(t_{2}, t_{1}\right)+\sum_{l_{\text {mm}}=1}^{L_{\text {mm}}}c_{l_{\text {mm}}}(\mathbf{r})v_{l_{\text {mm}}}\left(t_{2},t_{1}\right),\end{aligned}$$
where $$$t_1$$$ and $$$t_2$$$ denote the TE and FID dimensions, respectively. $$$\{v_{l_{x}}(t_2,t_1)\}$$$ denote basis (incorporating data-driven lineshape-adaption) spanning the subspaces for different components (water, lipid, metabolite and macromolecules), with component-specific orders $$$l_x$$$. Slab-specific nuisance signal removal was performed using the strategy described in Ref.[14]. For spatiospectral reconstruction, a subspace-constrained reconstruction was performed. More specifically, the multi-TE metabolite and macromolecule subspaces were combined in this step. A combined subspace was first learned using training data generated by a physics-driven strategy incorporating empirical distributions of spectral parameters¹⁵. The learned subspace was then adapted to experimental lineshape variations using spatial-dependent FIR filters (applied to low-resolution, the high-SNR counterpart of the data) and used in the reconstruction.

Multi-TE spectral quantification
After reconstruction, parameter estimation was performed in a task-specific fashion. Specifically, data from [65,80]ms were used for quantifying the metabolite and neurotransmitter (GABA and Glx) components using strategies developed in Ref.[12]. Data from [35,80,245]ms were used for estimating metabolite T2s using a multi-TE UoSS-based signal separation followed by T2 fitting strategy¹³. The macromolecule component was removed using a ProFit-based strategy¹⁷ to improve T2 estimation. More details on quantification are omitted due to the space limit.

Results

In vivo experiments were conducted on a 3T Prisma scanner using a 20-channel head coil (IRB approved), with 220x220x100 (head-foot direction) mm³ FOV, covering almost the entire brain. Figures 2-4 show some representative results from an experiment with 4 TEs and a nominal resolution of 3.4×3.4×5 mm³ (64x64 in-plane matrix size and 10 z-encodings for each slab). The total acquisition time is ~ 32 mins with 1.4s TR. High-quality spatially resolved multi-TE spectra from both slabs were produced (Fig. 2). High-resolution, high-SNR metabolite T2 maps across a large brain volume are shown in Fig. 3. Validation against a single-voxel multi-TE MRS (TE = [35,105,175,245]ms) showed excellent consistency for regional T2 values from the proposed method, but ~100-fold smaller voxels (Fig. 4). Metabolite and neurotransmitter maps across both slabs were shown in Fig. 5, which further demonstrate the impressive, multiparametric imaging capability enabled.

Conclusion

A novel accelerated J-resolved MRSI method was proposed, integrating SNR-efficient multi-slab excitation, optimized TEs selection, novel interleaved water acquisitions and subspace processing for the first time. Our method enabled high-resolution whole-brain metabolite, neurotransmitter and T2 mapping.

Acknowledgements

This work was supported in part by NSF-CBET-1944249 and NIH-NIBIB-1R21EB029076A