Info
Background:
Ultrashort-TE magnetic resonance spectroscopic imaging (MRSI) provides multiplexed molecular information with enhanced sensitivity. A key challenge is the presence of strong macromolecule (MM) signals (faster relaxation than small metabolites), which exhibit broad spectra strongly overlapping with metabolite peaks, complicating metabolite quantification [1]. The MM component also provides potentially valuable biological information [2]. Existing approaches to separate and quantify the MM signals either require additional long subject-specific MM acquisitions, more complex nonlinear parametric fitting, or insufficiently account for the spatial variability of MM signals [3-6]. A more robust and efficient approach to model and reconstruct the MM signals in ultrashort-TE MRSI data for enhanced simultaneous metabolite and MM mapping is desired.
Hypothesis/Objective:
The objective of this work is to develop a new approach for MM modeling and improve metabolite quantification in ultrashort-TE MRSI. This approach, initially developed and evaluated using rodent brain MRSI at ultrahigh fields (i.e., 9.4T), integrates optimized MM acquisition, data-driven prelearning of a low-dimensional MM subspace model, and subspace-constrained MRSI reconstruction. The proposed approach enables simultaneous MM and metabolite imaging, without requiring subject-specific MM acquisitions during routine ultrashort-TE scans while minimizing bias in signal models used in reconstruction. While this work focuses on preclinical rodent brain MRSI evaluation, our approach can be extended to human MRSI.
Methods:
MM acquisition and model learning: A double inversion recovery (DIR) and pulse-acquire ultrashort-TE chemical shift imaging (CSI) sequence was designed with recovery times numerically optimized for uniform brain metabolite suppression while maximally preserving MM signals at 9.4 T. Data quality was further enhanced using a custom field-inhomogeneity-corrected reconstruction and self-supervised denoising [7]. A low-dimensional subspace structure was identified from the ensemble of MM spectra from multiple mice and used as a learned representation for future experiments.
Joint MM and metabolite recon: The learned MM subspace was incorporated into a subspace reconstruction formulation for general ultrashort-TE data acquired without DIR. The signal is modeled as:
where denotes the overall MRSI signal, span the metabolite and MM subspaces (learned using the acquisition mentioned above), contain the unknown spatial coefficients, and represent the noise. With prelearned from training data, spatiospectral reconstruction was formulated as a spatially-regularized, subspace-constrained least-squares estimation of [8-9]. Simulations and in vivo studies were performed to evaluate the proposed method, in terms of improving metabolite mapping and simultaneous MM mapping.
Results:
Our optimized DIR-CSI sequence produced high-quality spatially resolved MM spectra in vivo (Fig. 1A). The ensemble of MM spectra from multiple mouse brains exhibited a low-rank behavior (Fig. 1B). A subspace was prelearned and then incorporated into the MRSI reconstruction. Results from simulations showed effective signal separation, improved metabolite reconstruction accuracy, and reduced fitting residuals compared to baseline approaches. Improved metabolite mapping over existing methods with simplified MM treatments was observed in vivo (Fig. 1C). Simultaneous metabolite and MM mapping were achieved at a 0.67 mm × 0.67 mm in-plane resolution using the prelearned MM subspace without requiring additional subject-specific MM acquisitions (Fig. 1C).
Conclusion:
A new MM modeling and separation method for ultrashort-TE MRSI at ultrahigh fields is developed and evaluated in mouse brains at 9.4 T. By integrating optimized DIR-based MM acquisition, data-driven MM subspace modeling, and augmented subspace-based reconstruction, the proposed method provides a practical and effective solution for explicit MM characterization and improved metabolite quantification in high-resolution preclinical MRSI.
Image/Figure:
Click to View
Image/Figure Caption:
Simultaneous in vivo metabolite and mm from our method.
(A) The proposed DIR-CSI acquisition produced high-quality voxel MM spectra in vivo.
(B) MM spectra from multiple mice can be well approximated by a low-dimensional subspace, supported by the rapid singular value decay of the Casorati matrix formed by concatenating all the voxel spectra (5% error at a rank-7 truncation, which was used in the reconstruction).
(C) Multimolecular maps estimated from an ultrashort-TE CSI data of a new mouse; Maps of creatine (Cre) and N-Acetylaspartic acid (NAA) are shown along with the corresponding estimated Cramér–Rao lower bound (CRLB) maps (indicating estimation uncertainty; lower is better). MM maps are shown in the last column. The proposed method (Bottom) was compared to a standard estimation approach (Top), i.e., a joint parametric fitting of metabolites and MMs (LCModel). Mean CRLB across brain voxels shows an apparent reduction of estimation uncertainty achieved by the proposed method, which also produced spatial distributions with better-defined tissue contrast.
Author
Suhao Liu
University of Illinois, Urbana-Champaign