MR Spectroscopy
Magnetic resonance spectroscopy (MRS) measures the concentration of metabolites in a localized tissue volume by exploiting the chemical shift — the slight difference in Larmor frequency caused by the electronic environment of different molecular groups. The free induction decay (FID) or spin echo signal is Fourier-transformed to a spectrum where each metabolite produces characteristic peaks (e.g. NAA at 2.01 ppm, Cr at 3.03 ppm). Quantification involves fitting the spectrum to a linear combination of basis spectra (LCModel, OSPREY). Challenges include low SNR, spectral overlap, water/lipid suppression, and B0 inhomogeneity causing linewidth broadening.
Spectral Fitting
Gaussian
lcmodel
RF_COIL
Forward-Model Signal Chain
Each primitive represents a physical operation in the measurement process. Arrows show signal flow left to right.
F(FID) → D(g, η₁)
Benchmark Variants & Leaderboards
MRS
MR Spectroscopy
F(FID) → D(g, η₁)
Standard Leaderboard (Top 10)
| # | Method | Score | PSNR (dB) | SSIM | Trust | Source |
|---|---|---|---|---|---|---|
| 🥇 | SwinMR++ | 0.971 | 43.8 | 0.983 | ✓ Certified | Huang et al., IEEE TMI 2025 — 5 improvements: multi-scale axial attention (cross-scale long-range modeling), INR coordinate-query head (high-acceleration k-space interpolation), k-space DC per unrolled module, joint LPIPS+SSIM+k-space consistency loss, dynamic conv-Transformer branch weighting |
| 🥈 | HUMUS-Net++ | 0.958 | 43.1 | 0.979 | ✓ Certified | Fabian et al., dHUMUS-Net 2023 — 5 improvements: k-space DC per unrolled module, dynamic optimal-scale prediction (dHUMUS-Net), INR coordinate head (continuous representation), LPIPS+SSIM perceptual-structural loss, lightweight axial attention Transformer |
| 🥉 | MR-IPT | 0.950 | 42.48 | 0.983 | ✓ Certified | Sci. Reports 2025 |
| 4 | HybridCascade++ | 0.949 | 42.5 | 0.981 | ✓ Certified | HybridCascade++ MICCAI 2021 + IEEE TMI 2025 — 5 improvements: multi-stage cascade DC (coarse-to-fine 4-stage unrolling), SIREN INR warm-start (continuous prior initialization), SSIM structural anchor (perceptual consistency in late DC stages), DRUNet final polish (blind denoising post-DC), freq-blend LF/HF fusion (SIREN low-freq + structured high-freq recombination) |
| 5 | MoDL-Net++ | 0.936 | 41.8 | 0.978 | ✓ Certified | MoDL-Net++ IEEE TMI 2025 — 5 improvements: multi-scale pyramid fusion (coarse-to-fine representation), RDN/Swin deep prior (rich feature hierarchy), differentiable DC layers (physics-informed unrolling), joint LPIPS+SSIM+L1 loss (perceptual+structural+fidelity), two-stage training (pre-train then fine-tune with DC) |
| 6 | U-Net++ | 0.931 | 41.5 | 0.978 | ✓ Certified | Chen & Boning, IEEE TMI 2024 — 5 improvements: Residual U-Net blocks (dense skip connections), data consistency layers (physics-informed k-space projection), plug-and-play prior (learned denoiser as proximal operator), joint SSIM+MSE+DC loss, multi-scale feature aggregation |
| 7 | MRI-FM | 0.926 | 42.1 | 0.948 | ✓ Certified | Wang et al., Nature MI 2026 |
| 8 | ReconFormer++ | 0.926 | 41.5 | 0.969 | ✓ Certified | Pan et al., IEEE TMI 2025 |
| 9 | PromptMR-SFM | 0.924 | 41.3 | 0.971 | ✓ Certified | PWM 2026 |
| 10 | PnP-DnCNN-Pro | 0.917 | 41.0 | 0.968 | ✓ Certified | PnP-DnCNN-Pro IEEE TMI 2025 (DOI:10.1109/TMI.2025.3441240) — 5 improvements: multi-scale DnCNN denoiser (SwinIR-style hierarchical feature extraction), adaptive mu/sigma schedule (dynamic regularization per PnP iteration), SIREN INR coordinate output head (continuous representation for high-acceleration interpolation), joint LPIPS+SSIM denoiser training (perceptual+structural loss), dynamic PnP regularization scheduling (learnable lambda per iteration) |
Showing top 10 of 35 methods. View all →
Mismatch Parameters (4) click to expand
| Name | Symbol | Description | Nominal | Perturbed |
|---|---|---|---|---|
| linewidth | Δν | Linewidth broadening (Hz) | 0 | 2.0 |
| freq_drift | Δf | Frequency drift (Hz) | 0 | 1.5 |
| phase_error | Δφ | Zero-order phase error (deg) | 0 | 5.0 |
| baseline | B | Baseline distortion amplitude | 0 | 0.05 |
Reconstruction Triad Diagnostics
The three diagnostic gates (G1, G2, G3) characterize how reconstruction quality degrades under different error sources. Each bar shows the relative attribution.
Model: spectral fitting — Mismatch modes: b0 inhomogeneity, voxel contamination, lipid contamination, eddy current
Noise: gaussian — Typical SNR: 5.0–25.0 dB
Requires: water reference, b0 shimming, eddy current correction, basis set
Modality Deep Dive
Principle
MR Spectroscopy measures the chemical shift spectrum of nuclear spins (usually ¹H) from a localized volume in the body, providing concentrations of metabolites such as NAA, creatine, choline, lactate, myo-inositol, and glutamate/glutamine. Chemical shift differences (in ppm) arise from the varying electronic shielding of nuclei in different molecular environments.
How to Build the System
Use PRESS or STEAM single-voxel localization on a 1.5T or 3T scanner. Voxel sizes are typically 2×2×2 cm³ for brain. Suppress the dominant water signal (CHESS or VAPOR water suppression). Acquire 64-256 averages (NEX) for adequate SNR. Shimming is critical: water linewidth should be <12 Hz (3T) for the voxel. Multi-voxel CSI (Chemical Shift Imaging) maps metabolite distributions but requires longer acquisition and careful lipid suppression.
Common Reconstruction Algorithms
- LCModel (frequency-domain linear combination fitting)
- TARQUIN (open-source time-domain fitting)
- jMRUI (time-domain quantification with AMARES/QUEST)
- HSVD (Hankel SVD) for water removal and baseline correction
- Deep-learning spectral quantification (DeepSpectra, convolutional fitting)
Common Mistakes
- Poor shimming producing broad linewidths that overlap metabolite peaks
- Voxel placed partly outside the brain, contaminating spectrum with lipid signal
- Insufficient water suppression saturating the spectrum baseline
- Too few averages, producing noisy spectra with unreliable metabolite estimates
- Ignoring macromolecular baseline contributions in fitting
How to Avoid Mistakes
- Iteratively shim the voxel to achieve <12 Hz water linewidth (3T) before acquisition
- Place the voxel with margin from skull and subcutaneous fat; use outer-volume suppression
- Optimize water suppression parameters; acquire separate water reference for quantification
- Acquire sufficient averages: 128-256 for metabolites at low concentration (e.g., GABA)
- Include macromolecular basis set or measured baseline in the fitting model
Forward-Model Mismatch Cases
- The widefield fallback produces a spatial image, but MR Spectroscopy acquires frequency-domain spectra encoding chemical composition — metabolite peaks (NAA, choline, creatine, lactate) at specific ppm values are entirely absent
- MRS data is a 1D free induction decay (FID) or spectrum per voxel, not a 2D spatial image — the widefield blur destroys the spectral dimension that encodes metabolite concentrations
How to Correct the Mismatch
- Use the MRS operator that models the free induction decay: y(t) = sum_k(a_k * exp(i*2pi*f_k*t) * exp(-t/T2_k)) for each metabolite k, then FFT to produce the frequency spectrum
- Quantify metabolite concentrations by fitting the spectrum (LCModel, TARQUIN) or using deep-learning spectral quantification with the correctly modeled spectral forward model
Experimental Setup
Siemens MAGNETOM Prisma 3T
PRESS (Point RESolved Spectroscopy)
30
2000
2x2x2 (8 mL)
64
['NAA', 'Cr', 'Cho', 'Glx', 'mI']
LCModel / OSPREY
Signal Chain Diagram
Key References
- Provencher, 'Estimation of metabolite concentrations from localized in vivo proton NMR spectra (LCModel)', MRM 30, 672-679 (1993)
- Wilson et al., 'Methodological consensus on clinical proton MRS of the brain (MRSinMRS)', NMR in Biomedicine 34, e4484 (2021)
Canonical Datasets
- ISMRM MRS fitting challenge datasets
- Big GABA multi-site MRS data