Deep learning water-unsuppressed MRSI at ultra-high field for simultaneous quantitative metabolic, susceptibility and myelin water imaging

Purpose: Magnetic Resonance Spectroscopic Imaging (MRSI) maps endogenous brain metabolism while suppressing the overwhelming water signal. Water-unsuppressed MRSI (wu-MRSI) allows simultaneous imaging of water and metabolites, but large water sidebands cause challenges for metabolic fitting. We developed an end-to-end deep-learning pipeline to overcome these challenges at ultra-high field. Methods:Fast high-resolution wu-MRSI was acquired at 7T with non-cartesian ECCENTRIC sampling and ultra-short echo time. A water and lipid removal network (WALINET+) was developed to remove lipids, water signal, and sidebands. MRSI reconstruction was performed by DeepER and a physics-informed network for metabolite fitting. Water signal was used for absolute metabolite quantification, quantitative susceptibility mapping (QSM), and myelin water fraction imaging (MWF). Results: WALINET+ provided the lowest NRMSE (< 2%) in simulations and in vivo the smallest bias (< 20%) and limits-of-agreement (+-63%) between wu-MRSI and ws-MRSI scans. Several metabolites such as creatine and glutamate showed higher SNR in wu-MRSI. QSM and MWF obtained from wu-MRSI and GRE showed good agreement with 0 ppm/5.5% bias and +-0.05 ppm/ +- 12.75% limits-of-agreement. Conclusion: High-quality metabolic, QSM, and MWF mapping of the human brain can be obtained simultaneously by ECCENTRIC wu-MRSI at 7T with 2 mm isotropic resolution in 12 min. WALINET+ robustly removes water sidebands while preserving metabolite signal, eliminating the need for water suppression and separate water acquisitions.

💡 Research Summary

**
This paper presents an end‑to‑end deep‑learning framework for water‑unsuppressed magnetic resonance spectroscopic imaging (wu‑MRSI) at ultra‑high field (7 Tesla) that simultaneously yields quantitative metabolic maps, quantitative susceptibility mapping (QSM), and myelin water fraction (MWF) images. Conventional MRSI relies on water suppression techniques (e.g., WET, VAPOR) to reduce the dominant water signal, which increases SAR, prolongs repetition time, and discards valuable information contained in the water resonance. Moreover, separate water acquisitions are required for absolute metabolite quantification, QSM, and MWF, leading to longer scan times and potential mis‑registration due to motion.

To overcome these limitations, the authors combined three technical innovations: (1) a non‑Cartesian ECCENTRIC k‑t sampling trajectory with ultra‑short echo time (TE = 0.9 ms) and optimized gradient spoiler waveforms that dramatically reduce mechanical vibrations and consequently water sideband amplitudes; (2) a novel deep‑learning network called WALINET+ that removes the main water peak, water sidebands, and lipid signals in a single step; and (3) a physics‑informed reconstruction network, DeepER, which recovers high‑quality images from compressed‑sensing accelerated data.

WALINET+ extends the previously published WALINET architecture by adding a second encoder and a Y‑net decoder that predict the nuisance spectrum (water, sidebands, lipids) from the raw complex spectrum. The predicted nuisance is subtracted, leaving only the metabolite component. Training employed a massive dataset of three million synthetic and experimentally derived spectra, including simulated metabolites, real water peaks extracted by HLSVD, sideband patterns measured in water phantoms, and in‑vivo lipid spectra. Data augmentation (frequency shifts, scaling, random phase) and L2 lipid projection were used to improve robustness. The network was trained for 4000 epochs on an NVIDIA A40 GPU using mean‑squared error loss and the Adam optimizer.

DeepER performs joint k‑space and temporal reconstruction, enabling an acceleration factor of 2 (and retrospectively up to 4) while preserving spectral fidelity. After reconstruction, the retained water signal serves three purposes: (i) it acts as an internal reference for absolute metabolite concentration; (ii) its phase information is used for QSM; and (iii) multi‑echo GRE data acquired at the same resolution provide ground‑truth QSM and MWF for validation.

In vivo experiments on healthy volunteers and a small patient cohort demonstrated that WALINET+ achieved a normalized root‑mean‑square error (NRMSE) below 2 % in simulations and a bias under 20 % with limits of agreement ±63 % when compared to conventional water‑suppressed MRSI. Metabolites such as creatine and glutamate showed 10‑20 % higher signal‑to‑noise ratio in the water‑unsuppressed acquisition. QSM maps derived from wu‑MRSI agreed with GRE‑based QSM (mean bias 0 ppm, limits ±0.05 ppm), and MWF estimates matched GRE‑based MWF (mean bias 5.5 %, limits ±12.75 %). The entire protocol produced whole‑brain coverage at 2 mm isotropic resolution in approximately 12 minutes, a substantial time saving compared with separate acquisitions.

The study demonstrates that water‑unsuppressed MRSI, when coupled with advanced deep‑learning preprocessing and reconstruction, can provide high‑quality metabolic, susceptibility, and myelin water information in a single, time‑efficient scan. This multiparametric approach is particularly advantageous for neurodegenerative and neuroinflammatory diseases where combined assessment of metabolites, iron deposition, and myelin integrity is clinically relevant. Limitations include the need for 7 T hardware to achieve the reported SNR and resolution, and the relatively modest size of the training cohort, which may affect generalization to other scanners, field strengths, or patient populations. Future work should explore transfer learning to lower field strengths, larger heterogeneous datasets, and integration with clinical decision‑support pipelines.