MRS Data Analysis and Quality Control

1) Preprocessing: everything done to the raw data to prepare it for fitting (combine coils, align and average, correct artefacts, remove water).
2) Fitting: estimate each metabolite's signal (arbitrary units) with a model, usually a linear combination of basis spectra.
3) Quantification: convert those numbers into reportable units (mmol/kg wet weight, or ratios).
Quality control runs through all three.

A good fitting algorithm cannot rescue bad data: motion, poor shim, poor water suppression, lipid contamination, eddy currents or chemical-shift artefacts all give unreliable concentrations.
Good data fitted with the wrong settings are also unreliable. Fitting software is automatic but not push-button: it must be adapted to the sequence, TE, B0 and nucleus.

1) Frequency → peak position.
2) Amplitude → peak area.
3) Damping rate (1/T2*) → linewidth (FWHM = damping/π).
4) Phase at t = 0 → zero-order phase.
5) Beginning time (delay before recording) → first-order phase.

Zero-order: the same phase error on every peak (from the receiver/coil phase); all peaks mix absorption and dispersion alike.
First-order: the error grows linearly with frequency. Cause: recording starts a delay t0 after the pulse, so each component has already turned by 2πf·t0 (faster components further).
Both are corrected in preprocessing or included in the fit.

No. They are necessary but not sufficient: two spectra can have the same SNR and FWHM while one is ruined by residual water, lipid contamination or a spurious echo.
Quality control means looking for artefacts, then correcting or rejecting the data.

Sample and protocol: voxel volume (signal ∝ volume), number of averages (√N), sequence (e.g. STEAM keeps half), TE and TR.
Adjustments: RF power (flip angle), coil tuning/matching, B1 shimming, slice rephasing, gradient delays.
Hardware: receiver noise figure, the coil (e.g. a cryoprobe).
Small peaks vanish first as SNR drops.

Broadening: hardware performance, subject motion, and averaging shots whose frequency has drifted (a sum of shifted peaks is a wider peak).
Distorted (asymmetric) shape: a misadjusted shim term (e.g. Z2) or eddy currents. A bad shape misleads the fit even if the width looks fine.

Causes: type of WS and its settings (RF power, delays), sensitivity to T1 and B1, broad water from a bad shim (spills outside the suppression band), water from outside the voxel.
Appearance: large residual water at 4.7 ppm, distorted baseline and amplitude/phase distortions of nearby peaks.

Fat outside the voxel (scalp, muscle) leaks in through imperfect localization or slice profiles, weak crushers, insufficient OVS, CSDE or motion → broad peaks at 0.9–1.3 ppm that swamp lactate, alanine and macromolecules.
Remedies: OVS bands, double localization (OVS + sequence, or LASER), stronger crushers, high-bandwidth pulses.

Water far from the voxel sits in a very different field (strong 2nd-order shims make distant fields far off), so it is off-resonance, escapes water suppression, and may not be fully crushed. It refocuses late as a delayed echo inside the FID.
In the spectrum: a spurious ringing feature. Remedies: stronger crusher gradients, OVS.

Signal: SNR > 3 for major resonances; adequate linewidth; symmetric lineshape; good water suppression; no lipid contamination; no other visible artefacts; watch for hidden ones (CSDE, localization errors, signal cancellation).
Fit: acceptable CRLBs; residuals without unexplained features.

Preclinical: anaesthetised animal, little motion, time to shim/WS/OVS, MRS experts at the scanner, moderate data.
Clinical: awake patients, motion, little time, often non-expert operators, huge amounts of data. Artefacts slip through, and standard software without QC can report false changes (e.g. apparent Cho/Cr changes in a glioma) → wrong decisions.

Fitting models assume a clean, ideal signal and do not account for motion, frequency drift, eddy currents or misaligned shots, so these would bias the fit.
Corrections should be automated and built into the analysis software.

1) Combine RF coil channels.
2) Remove motion-corrupted shots.
3) Correct frequency and phase drift (align shots).
4) Combine (average) shots.
5) Eddy-current correction.
6) Remove residual water (HSVD).
7) Phase correction, then fit.
(Zero-filling and apodization: display only.)

B0 drifts over a long scan and motion changes phase. Averaging misaligned shots broadens the peaks and loses signal.
For add/subtract methods (ISIS, SPECIAL), misalignment also leaves signal from outside the voxel after subtraction.

Acquire the unsuppressed water signal from the same voxel with the same sequence. Its time-varying phase reflects the eddy-current field; subtract that phase, point by point, from the metabolite FID → clean lineshape.

For display only, not before fitting.
Zero-filling pads the FID with zeros: a finer frequency grid, no new information. Apodization (line broadening) multiplies the FID by a decaying window: less noise, broader peaks. Same trade-offs as in k-space for imaging.

Complexity: many resonances, heavy overlap, macromolecules and residual water underneath.
Unpredictable lineshape: residual eddy currents, field inhomogeneity.
Unpredictable baseline: macromolecules, lipids, partly suppressed water.

Peak fitting: fit each peak with a Lorentzian/Gaussian/Voigt shape; fine for few well-separated peaks (X-nuclei), fails with overlap.
HLSVD: splits the FID into damped sinusoids with no chemical prior knowledge; fast, used for water removal and to quantify simple spectra (jMRUI).
Basis-set (linear combination): spectrum = Σ c_m × basis_m + macromolecules + baseline; the standard for overlapped 1H spectra.

One spectrum per metabolite for exactly the sequence, TE and B0 used, either measured (metabolite solutions) or simulated quantum-mechanically from known chemical shifts and J-couplings (NMRScopeB, Vespa, GAMMA, GAVA).
Plus a macromolecule component (measured with metabolite-nulling inversion, since MM T1 is short, or modelled) and lipids where relevant (e.g. tumours).

When their patterns overlap too much to separate reliably, the fit reports the combination: NAA + NAAG, Glu + Gln (Glx), GPC + PCho (tCho), Cr + PCr (tCr).

Vespa, TARQUIN and jMRUI are free; LCModel is listed as paid on the slide. All accept all data types and include preprocessing.
Lineshape: TARQUIN uses Voigt, jMRUI Lorentzian, LCModel estimates it from the data. LCModel needs a supplied basis set and models the baseline with splines.

The theoretical minimum uncertainty (standard deviation) of a fitted amplitude, given the noise and the model, usually quoted as % of the estimate.
It assumes the model is perfect, so it says nothing about whether the model or fit is good. Packages compute it differently, and it does not replace real repeatability studies.

A low concentration gives a large %CRLB even when the absolute uncertainty is fine. Rejecting values above a %CRLB threshold removes low estimates preferentially → the group mean is biased upwards.
Better: compare with the typical CRLB in a normal cohort.

Residual = data − fit; it should look like pure noise. Structure in it means something is missing from the model (a metabolite, the right lineshape or baseline, an artefact).
FQN (fit quality number) = residual variance / noise variance: ≈ 1 means the fit explains everything but noise.

1) External reference: phantom of known concentration (correct for coil loading and B1).
2) Internal ratio to tCr, NAA or Cho: simple, scale cancels, but the reference may itself change with disease.
3) Internal water reference: unsuppressed water from the same voxel, corrected for T1, T2, water content (and small MT effects). Units: mmol/kg wet weight.

[M] = (S_M/S_W) × (2/n_M) × [W] × (R_W/R_M), with R = (1 − e^(−TR/T1))·e^(−TE/T2) for each signal. 2 = protons in water, n_M = protons in the metabolite peak, [W] = 55.5 M × tissue water content.
Very short TE → T2 factors ≈ 1; fully relaxed long TR → T1 factors ≈ 1.

[W] = 0.8 × 55,500 ≈ 44,400 mmol/kg. NAA's singlet has 3 protons, water 2.
[NAA] = (1/3000) × (2/3) × 44,400 ≈ 9.9 mmol/kg.

Any MRS sequence plus phase-encoding gradients: a spectrum in every voxel of a grid → metabolite maps (often ratios like [Tau]/[tCr]).
No readout gradient (the readout records the spectrum). Each phase step adds a position-dependent phase to each voxel's whole spectrum: FT over time → ppm, FT over the steps (k) → position.
Advantage: many regions compared in one scan; also works for 31P, 2H, functional and diffusion-weighted MRSI.

1) Low metabolite concentrations → low SNR, long scans.
2) Shimming a whole slice: B0, linewidth and frequency vary voxel to voxel.
3) Water suppression over the whole organ: one suppression frequency, but water's frequency varies (inhomogeneity, drift) → uneven suppression.
4) Lipids leak in through the point-spread function.
5) Hardware limits and in-house sequences.

3 pulses (PRESS/STEAM/SPECIAL-MRSI): a box inside the brain, split into a grid → "square maps"; lipids avoided, but no full-brain coverage.
2 pulses (slice-selective spin echo): a slice.
1 pulse (FID-MRSI): a slice → whole-slice coverage, but strong scalp lipids need saturation bands.

The phase a gradient gives depends only on position, not on the metabolite's frequency, so all metabolites of a voxel land in the same place.
Trap: CSDE still exists along directions defined by slice-selective pulses (the slice, or the edges of a PRESS box).

T_MRSI = N_avg × Nx × Ny (× Nz) × TR; T_SVS = N_avg × TR (~5 min human, 10–15 min preclinical).
At TR 2 s: 32 × 32 = 34 min; 16 × 16 × 16 = 136 min.
Faster: short TR (low flip angle), several k-points per TR (spatial–spectral encoding), undersampling (parallel imaging, compressed sensing: AF 4, 119 → 28 min), prior-knowledge reconstruction.

Every excitation collects signal from the whole excited volume, and every phase-encode step contributes to every voxel.
So a 32 × 32 scan with one average acts like 1024 averages of each voxel. Compare methods with SNR per unit time = SNR/√(total time).

Saturation bands, more phase-encode steps, k-space weighting, lipid suppression in processing.
A Hamming filter applied afterwards cuts the PSF ripples but broadens the PSF (lower resolution) and down-weights high-k data you already paid for (less sensitivity per unit time). Acquisition weighting measures central k-space more often, shaping the PSF without that waste.

One slice-selective pulse, phase encoding, FID after an ultra-short acquisition delay (~1–1.3 ms); short TR (200–800 ms), low flip angle.
Pros: negligible J-evolution and T2 loss; fast (31 × 31, TR 811 ms ≈ 13 min at 14.1 T), so more phase steps → better PSF; whole-slice coverage.
Cons: the delay gives a first-order phase; whole-slice lipids need strong saturation.

Four metrics per voxel: SNR, linewidth (FWHM), B0 shift, CRLB.
Check the maps before fitting (no masking), then mask voxels that fail after fitting. MRS4Brain toolbox criteria: SNR above 75% of the mean, FWHM below 125% of the mean, CRLB < 40% (one slide example uses 30%).

A free EPFL toolbox for MRSI, single-voxel and diffusion-weighted spectroscopy.
MRSI: read Bruker data → brain mask (water signal) → HSVD water removal → lipid suppression → quick maps (linewidth, ΔB0, SNR = NAA height/σ noise) → LCModel fit → quality criteria → metabolite maps with atlas segmentation (SIGMA).