MRSuite/mlp-supervised-7t-slaser

A multi-layer perceptron that reads concentrations and signal parameters off a spectrum.

An MRSuite model, mrsuite.models.mlp.MLP. Paper: https://arxiv.org/abs/2511.23135

Use

import mrsuite.models

model = mrsuite.models.load("MRSuite/mlp-supervised-7t-slaser")
results = model(data)          # one spectrum as NIfTI-MRS, or a batch as NIfTI-MRS+

It runs as a step in any MRSuite pipeline, and each result's provenance records its exact weights (SHA-256 45a62ce6c27b96d48c1cd425a1ddd88e4ce7232ecf867395fb9c754ff6b142c4).

It takes single spectra acquired as its training data, and refuses others:

Acquisition
Bandwidth [Hz]	3000.0
Centre [ppm]	4.65
Spectrometer frequency [MHz]	298.200012
Points	1024

Training

{
  "model": "mrsuite.models.mlp.MLP",
  "checkpoint": {
    "path": "data/checkpoints/mlp_super/epoch=0-step=7537664.ckpt",
    "sha256": "7ef95bb4916e915ce5847cf17229f18ca15742e9cb87dddc5f9a0991dd9f829c"
  },
  "job": {
    "model": "mlp",
    "basis": "7T_sLASER_OIT_TE34.basis",
    "basisFmt": "7tslaser",
    "ppmlim": [
      0.5,
      4.0
    ],
    "dataType": "aumc2",
    "width": 512,
    "depth": 3
  },
  "converted with": {
    "ood-robust-mrs": "451cf71"
  },
  "run": {
    "id": "wyoo5p40",
    "settings": {
      "activation": "elu",
      "adaptMode": "per_spec",
      "arch": "mlp",
      "basisFmt": "7tslaser",
      "batch": 16,
      "callback": "val_conc_loss",
      "check_val_every_n_epoch": null,
      "checkpoint_path": "",
      "conv_depth": 3,
      "dataType": "aumc2",
      "decomModel": "unet",
      "depth": 3,
      "device": "cuda:0",
      "dropout": 0,
      "flow_args": {
        "beta": 0,
        "flow": "SNFVAE",
        "num_flows": 8
      },
      "gpu_selection": 0,
      "gradient_clip_val": 0,
      "initMethod": "nn",
      "innerEpochs": 100,
      "innerLoss": "mse_specs",
      "innerLr": 0.001,
      "kernel_size": 3,
      "load_model": false,
      "loss": "mae_all_scale",
      "lr": 0.0001,
      "max_epochs": -1,
      "max_steps": -1,
      "maxIter": 10,
      "model": "nn",
      "num_workers": 4,
      "online": true,
      "optimizer": "theseus",
      "path2basis": "../Data/BasisSets/7T_sLASER_OIT_TE34.basis",
      "path2vae": "",
      "ppmlim": [
        0.5,
        4
      ],
      "project": "ComparisonDD",
      "reg_l1": 0,
      "reg_l2": 0,
      "shuffle": true,
      "skip_train": false,
      "specType": "auto",
      "stepSize": 0.1,
      "stride": 1,
      "trueBatch": 16,
      "val_check_interval": 256,
      "val_size": 1024,
      "width": 512
    }
  }
}

Training data

How the training spectra were made, reported as the MRSsynMRS minimum standard for synthetic MRS data asks (LaMaster et al., arXiv 2602.23463). The full recipe is in recipe.json.

MRSsynMRS	Value
Experiment
ID	OoD-Robust-MRS-Quantification dataType 'aumc2' (aumcParams2, aumcConcs), runs wyoo5p40 (supervised) and 48ke5d6u (self-supervised)
Nucleus	1H
Software
Basis Set Simulation
Name	7T_sLASER_OIT_TE34.basis (LCModel)
Repository (Link, DOI, or RRID)	https://github.com/CIC-methods/FID-A (the simulator, per the basis header's source path 7T_FIDA_sLASER_OIT_TE34)
Signal Model
Name	VoigtModel (simulation/sigModels.py): each basis FID times exp(-(1jeps + gamma + sigma^2t)t), FFT, times its amplitude and exp(-1j(phi0 + phi1*f)), plus a polynomial baseline
Repository (Link, DOI, or RRID)	https://github.com/julianmer/OoD-Robust-MRS-Quantification (451cf71), arXiv:2511.23135
Additional Code
Name	simulation/simulation.py, simulation/dataModules.py (SynthDataModule.get_batch), utils/components.py; FSL-MRS reads the basis
Repository (Link, DOI, or RRID)	https://github.com/julianmer/OoD-Robust-MRS-Quantification (451cf71)
Purpose	draws the parameters and adds the noise and any random walk and random peaks; simulated online, fresh spectra for every training batch
Pulse Sequence
Field Strength [T]	7
Localization	sLASER (OIT pulses, per the basis name)
Spectral Width [Hz]	3000
Number of Points	1024
Voxel Size (X, Y, Z) [mm]	—
Sequence Timing(s)
Repetition Time(s) [ms]	—
Echo Time(s) [ms]	34
Mixing Time(s) [ms]	—
Inversion Time(s) [ms]	—
Water Suppression
Method	—
Bandwidth(s) [Hz]	—
Shimming Method	—
Radio Frequency Pulse(s)
Shape	—
Duration(s)	—
Delay(s)	—
Slice-Select Gradients
Shape	—
Duration(s)	—
Delay(s)	—
Crusher/Rephasing Gradients
Shape	—
Duration(s)	—
Delay(s)	—
Editing
Type of editing	none
Number of subexperiments	1
Editing bands per subexperiment	—
Metabolites
Population
Age Range	—
Sex	—
Healthy included	—
Clinical included	—
Region of interest	—
Distribution	uniform: each parameter drawn independently and uniformly over its [low, high] (numpy.random.uniform)
Component dropout probability	0
Units	a.u.: the amplitude multiplying each basis FID as FSL-MRS reads the LCModel .basis (inverse FFT, conjugated) and resamples it to 3000 Hz and 1024 points, not rescaled (each metabolite's function also holds a reference peak at 0 ppm, outside the model's window); not mM
Sources of Parameter Values
Concentration Ranges (Link or DOI)	per name the widest of LCModel and FSL-MRS (Newton) fits of the AUMC 7 T data and de Graaf 2019's literature ranges (simulation/simulationDefs.py: aumcConcs over aumcConcsLCM, aumcConcsNewton, normConcs); arXiv 2511.23135
T2 Relaxation Ranges (Link or DOI)	—
T1 Relaxation Ranges (Link or DOI)	—
Concentration Ranges
Ala	0 – 1.6
Asc	0 – 4.9
Asp	0 – 4.8
Cr	3.9 – 12.3
GABA	0 – 4
Gln	0 – 6.8
Glu	6 – 17.9
Gly	0 – 1
GPC	0 – 3.6
GSH	0 – 3.6
mIns	4 – 12.1
Lac	0 – 3.1
NAAG	0 – 2.5
NAA	7.5 – 16.3
PCh	0 – 2.4
PCr	0 – 5.5
PE	0 – 5.2
Scyllo	0 – 0.6
Ser	0 – 7.3
Tau	1.2 – 6
MM_CMR	0 – 400
T2 Relaxation Ranges	not per metabolite: one Lorentzian decay for all basis functions, exp(-gamma*t) with gamma in [2, 25] 1/s (1/gamma 40-500 ms, FWHM gamma/pi 0.6366-7.958 Hz), on top of each basis function's own linewidth
T1 Relaxation Ranges	not applied
Metabolite-Level Frequency Shifts	not applied: one global shift (see Artifacts)
Components
Number Spectral Averages	1
Number of Acquisitions per Excitation	1
Dynamic Experiment
Diffusion B-values	—
Diffusion Model (Name, DOI)	—
fMRS Target Metabolites and Percent Changes	—
Macromolecule
Available in Isolation	yes
Simulated, Modeled, Experimental, or Mixture	the basis function MM_CMR (header ID 7T_MM_CMRR, from MM_295.1_MHz_CMRR_4kHz.RAW), sampled like the metabolites; measured or modelled is not stated
Description	MM_CMR
Spectral Baseline - nonMMBG
Model Used (Equation, Description, or DOI)	second-order polynomial added to the frequency-domain spectrum, real and imaginary parts separately: coefficients of 1, x and x^2 (less its mean), x from -1 to 1 across the model's input window 0.50-3.99 ppm, zero outside it; a coefficient is its term's amplitude in a.u.
Available in Isolation	yes
Applied	yes
If Analytical, Parameter Ranges (added)
Re 1	-600 – 200
Im 1	-800 – 300
Re x	-1000 – 600
Im x	-600 – 1000
Re x^2	-1600 – 200
Im x^2	-400 – 1000
Noise
Available in Isolation	yes
Distribution	additive complex Gaussian in the frequency domain (on the unnormalised FFT spectrum), real and imaginary parts independent
Amplitude Scaling
std of each of the real and imaginary parts, a.u., uniform per spectrum	10 – 7071.07
B0 Map
B0 Map Model Used	—
Available in Isolation	—
Applied	no
Tissue Segmentation
Brain Map Model Used	—
Voxel Percentage(s)	—
Applied	no
Artifacts
Zero-Order Phase
Phase Range
rad, as exp(-1j*phi0)	-0.5 – 0.5
deg	-28.65 – 28.65
Applied	yes
First-Order Phase
Phase Range
rad/Hz, as exp(-1jphi1f)	-1e-05 – 1e-05
deg/ppm	-0.1709 – 0.1709
Amplitude Range	—
PPM Reference [ppm]	-0.38
Applied	yes
Frequency Shifts - Global
Frequency Shift Range
rad/s, as exp(-1jepst)	-10 – 10
Hz	-1.592 – 1.592
ppm	-0.005337 – 0.005337
Applied	yes
Residual Water
Model Used (Equation, Description, or DOI)	—
If Analytical, Parameter Ranges	—
Available in Isolation	—
Applied	no
Chemical Shift [ppm]	—
Number of Subcomponents	—
Amplitude Scaling [a.u.]	—
Unsuppressed Water
Model Used (Equation, Description, or DOI)	—
Available in Isolation	—
Applied	no
Lipid
Model Used (Equation, Description, or DOI)	—
Available in Isolation	—
Applied	no
Eddy Currents
Model Used (Equation, Description, or DOI)	—
If Analytical, Parameter Ranges	—
Available in Isolation	—
Applied	no
Out-of-Voxel
Model Used (Equation, Description, or DOI)	—
If Analytical, Parameter Ranges	—
Available in Isolation	—
Applied	no
Motion
Model Used (Equation, Description, or DOI)	—
If Analytical, Parameter Ranges	—
Available in Isolation	—
Applied	no
Data Quality
Signal Noise Ratio Range [a.u.]
Range	8.9 – 1384.8
Metabolite-level SNRs recorded	no
Definition of SNR used	maximum of the real part of the noise-free spectrum (no phases, baseline or artifacts) in 1.9-2.1 ppm, the NAA singlet, over the standard deviation of the noise's real part; 1st-99th percentile of 2000 spectra drawn with the original code (numpy seed 0)
Linewidths
Metabolite-level Linewidths recorded	no, per spectrum: gamma and sigma (1/s) in the returned parameters; they add FWHM gamma/pi 0.6366-7.958 Hz (Lorentzian) and 2*sqrt(ln2)*sigma/pi 1.06-13.25 Hz (Gaussian)
Reference Metabolite and Moiety	NAA, CH3 singlet (2.01 ppm)
Reference Metabolite Linewidth Range [Hz]	3.2 – 19.7

Configuration

{
  "acquisition": {
    "bandwidth": 3000.0,
    "centre": 4.65,
    "frequency": 298.200012,
    "points": 1024
  },
  "depth": 3,
  "first": 602,
  "last": 957,
  "metabolites": [
    "Ala",
    "Asc",
    "Asp",
    "Cr",
    "GABA",
    "Gln",
    "Glu",
    "Gly",
    "GPC",
    "GSH",
    "mIns",
    "Lac",
    "NAAG",
    "NAA",
    "PCh",
    "PCr",
    "PE",
    "Scyllo",
    "Ser",
    "Tau",
    "MM_CMR"
  ],
  "model": "mrsuite.models.mlp.MLP",
  "width": 512
}