WALINET: A water and lipid identification convolutional neural network for nuisance signal removal in 1 H $$ {}^1\mathrm{H} $$ MR spectroscopic imaging

Abstract

Purpose: Proton magnetic resonance spectroscopic imaging ( ${}^{1}H$ -MRSI) provides noninvasive spectral-spatial mapping of metabolism. However, long-standing problems in whole-brain ${}^{1}H$ -MRSI are spectral overlap of metabolite peaks with large lipid signal from scalp, and overwhelming water signal that distorts spectra. Fast and effective methods are needed for high-resolution ${}^{1}H$ -MRSI to accurately remove lipid and water signals while preserving the metabolite signal. The potential of supervised neural networks for this task remains unexplored, despite their success for other MRSI processing.

Methods: We introduce a deep learning method based on a modified Y-NET network for water and lipid removal in whole-brain ${}^{1}H$ -MRSI. The WALINET (WAter and LIpid neural NETwork) was compared with conventional methods such as the state-of-the-art lipid L2 regularization and Hankel-Lanczos singular value decomposition (HLSVD) water suppression. Methods were evaluated on simulated models and in vivo whole-brain MRSI using NMRSE, SNR, CRLB, and FWHM metrics.

Results: WALINET is significantly faster and needs 8s for high-resolution whole-brain MRSI, compared with 42min for conventional HLSVD+L2. WALINET suppresses lipid and water in the brain by 25-45 and 34-53-fold, respectively. WALINET has better performance than HLSVD+L2, providing: (1) more lipid removal with 41% lower NRMSE; (2) better metabolite signal preservation with 71% lower NRMSE in simulated data; 155% higher SNR and 50% lower CRLB in in vivo data. Metabolic maps obtained by WALINET in healthy subjects and patients show better gray-/white-matter contrast with more visible structural details.

Conclusions: WALINET has superior performance for nuisance signal removal and metabolite quantification on whole-brain ${}^{1}H$ -MRSI compared with conventional state-of-the-art techniques. This represents a new application of deep learning for MRSI processing, with potential for automated high-throughput workflow.

Keywords: brain; metabolite quantification; mr spectroscopic imaging; ultrahigh‐field mr; water and lipid removal.

FIGURE 1

WAter and LIpid neural NETwork (WALINET): Top: Y‐Net architecture containing four convolutional blocks in each encoder and decoder, followed by a MaxPooling or Upsampling layer. Additional convolutional blocks are incorporated in the bottleneck and as a final layer. A lipid projection operator at the beginning of the second encoder enhances the distinguishability between metabolites and lipids. Bottom: The WALINET is embedded into the MRSI processing pipeline shown at the bottom, which includes Fourier transformation, B0 correction, low‐rank model, and spectral quantification.

FIGURE 2

Simulation results. (A) Comparison of lipid suppression by LIPNET and L2. Input spectra contaminated by lipids are shown in the first row (black), second row shows metabolite spectra recovered by L2 regularization (orange), and third row shows the metabolite spectra recovered by LIPNET (red). (B) Input spectra contaminated by water and lipids are shown on the first row (black), in the second row metabolite spectra processed with HLSVD+L2 are displayed, and the metabolite spectra predicted by WALINET (red) are plotted below. The ground‐truth metabolite spectra (blue) are overlaid in all spectral plots. Normalized root‐mean‐squared error (NRMSE) is computed for the whole spectrum (9.0–0.0 ppm), the metabolic range (4.2–1.9 ppm), and the lipid range (1.9–0.7 ppm) for each method LIPNET or L2 in (A) and WALINET or HLSVD+L2 in (B). Separate evaluation of specific spectral ranges allows an individual assessment of the preservation of metabolic signals, as well as the effectiveness of lipid and water suppression. Arrows indicate the position of the main peaks of water, lipids, and metabolites.

FIGURE 3

Comparison of lipid removal on in vivo 2D MRSI by LIPNET and L2 for three values of the regularization parameter $\beta$, the optimal value ($3.69\ast 10^6$), double the optimal value, and zero for no lipid suppression. Metabolic maps are shown for NAA+NAAG, the corresponding CLRB, SNR computed by LCModel, and residual lipid signal. Spectra from several brain voxels are shown for each method, the white trace shows the measured spectrum, and the red trace shows LCModel fit.

FIGURE 4

Comparison of water removal methods on in vivo 2D MRSI, including WALINET, HLSVD+LIPNET, LIPNET and L2 without water suppression. WET water suppression was used during acquisition. Metabolic maps and the corresponding CRLBs are shown for NAA+NAAG, Cr+PCr, and Glu together with maps of the residual water signal and examples of spectra (voxel locations are indicated by arrows, red line indicates LCModel fit and the white line indicates the experimental spectra).

FIGURE 5

Comparison of combined water & lipid removal on in vivo 3D MRSI. Results are shown for WALINET, HLSVD+LIPNET, and HLSVD+L2, including metabolic maps for NAA, Glutamate, Inositol, residual lipid signal, residual water signal, and SNR computed by LCModel. Selected spectra from individual voxels indicated by white arrows on the anatomical images are shown at the bottom (white trace shows measured spectrum, red trace shows LCModel fit).

FIGURE 6

Boxplots of spectral quality metrics obtained be WALINET, HLSVD+LIPNET, and HLSVD+L2 in 2D and 3D in vivo MRSI data.