Hugging Face's logo

MRSuite
/
wand-ismrm2016

Safetensors
mrsuite
model_hub_mixin
mrs
pytorch_model_hub_mixin
Model card Files
xet
 Community

MRSuite/wand-ismrm2016

Decomposes a spectrum into its metabolites, baseline and artifacts in the wavelet domain.

An MRSuite model, mrsuite.models.wand.WAND. Paper: https://doi.org/10.1002/nbm.70038

Use 

import mrsuite.models

model = mrsuite.models.load("MRSuite/wand-ismrm2016")
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 635e74f89a9b6a33ea4fc757005f84c3986284d86f2e00334a111fd125f41107).

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

Acquisition
Bandwidth [Hz] 4000.0
Centre [ppm] 4.65
Spectrometer frequency [MHz] 123.22
Points 2048

Training 

{
  "model": "mrsuite.models.wand.WAND",
  "checkpoint": {
    "path": "data/checkpoints/wand/epoch=0-step=1195264.ckpt",
    "sha256": "182bf931ba58485465f5296aa31ef9a59bd84f0f01b23b4b8002239a7a87d642"
  },
  "job": {
    "model": "wand",
    "basis": "data/2016_fitting_challenge/basisset_JMRUI",
    "specType": "synth",
    "arch": "unet",
    "dataType": "norm_rw_p"
  },
  "converted with": {
    "wand-for-mrs": "2642659"
  },
  "run": {
    "id": "sskt2pvt",
    "settings": {
      "arch": "unet",
      "basisFmt": "",
      "batch": 16,
      "check_val_every_n_epoch": null,
      "dataType": "norm_rw_p",
      "device": "cuda:3",
      "dropout": 0,
      "gpu_selection": 3,
      "l1_reg": 0,
      "l2_reg": 0,
      "learning": 0.0001,
      "load_model": false,
      "loss": "mse",
      "max_epochs": -1,
      "max_steps": -1,
      "model": "nn",
      "num_workers": 4,
      "path2basis": "../Data/BasisSets/ISMRM_challenge/basisset_JMRUI/",
      "path2trained": "",
      "shuffle": true,
      "skip_train": false,
      "specType": "synth",
      "trueBatch": 16,
      "val_check_interval": 256,
      "val_size": 1024
    },
    "W&B (added)": {
      "started": "2024-07-10T15:48:49.107867",
      "host": "EEMATLAB-OP02",
      "program": "sweep.py",
      "stopped at step": 1271899,
      "runtime [s]": 676609,
      "checkpoint": "step 1195264, the lowest validation loss (ModelCheckpoint(monitor='val_loss'), train.py)"
    }
  }
}

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 WAND-for-MRS dataType 'norm_rw_p' (normParamsRWP, normConcs), run sskt2pvt
Nucleus 1H
Note (added) the training-time code (forStudents/Code, 2024-07-10) draws the data as WAND-for-MRS 2642659 does: the same normParamsRWP and normConcs ranges for every basis function, simulation.py identical, SynthDataModule and the random walk and peaks unchanged
Software
Basis Set Simulation
  Name ISMRM 2016 MRS fitting challenge basis set, jMRUI (basisset_JMRUI): density-matrix simulation in MATLAB with ideal pulses, PRESS TE 30 ms (TE1 11 ms, TE2 19 ms), from measured and published chemical shifts and J-couplings (MarjaΕ„ska et al., Magn Reson Med 2021, doi:10.1002/mrm.28942)
  Repository (Link, DOI, or RRID) https://github.com/wtclarke/mrs_fitting_challenge, doi:10.1002/mrm.28942
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/WAND-for-MRS (2642659), doi:10.1002/nbm.70038
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/WAND-for-MRS (2642659)
  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] 2.89
Localization PRESS (TE1 11 ms, TE2 19 ms)
Spectral Width [Hz] 4000
Number of Points 2048
Voxel Size (X, Y, Z) [mm] β€”
Sequence Timing(s)
  Repetition Time(s) [ms] TR >> T1 (no T1 weighting)
  Echo Time(s) [ms] 30
  Mixing Time(s) [ms] β€”
  Inversion Time(s) [ms] β€”
Water Suppression
  Method β€”
  Bandwidth(s) [Hz] β€”
Shimming Method β€”
Radio Frequency Pulse(s)
  Shape ideal
  Duration(s) β€”
  Delay(s) β€”
Slice-Select Gradients
  Shape β€”
  Duration(s) β€”
  Delay(s) β€”
Crusher/Rephasing Gradients
  Shape β€”
  Duration(s) β€”
  Delay(s) β€”
Editing
  Type of editing none
  Number of subexperiments 1
  Editing bands per subexperiment β€”
Metabolites
Population
  Age Range β€”
  Sex β€”
  Healthy included β€”
  Clinical included β€”
Region of interest β€”
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 jMRUI files, not rescaled (a singlet's |FID(0)| is its number of protons, plus 0.1 from a reference peak at 0 ppm); not mM
Sources of Parameter Values
  Concentration Ranges (Link or DOI) de Graaf, In Vivo NMR Spectroscopy, 3rd ed. (2019), per the original's comment on normConcs (simulation/simulationDefs.py); Mac's range assumed
  T2 Relaxation Ranges (Link or DOI) β€”
  T1 Relaxation Ranges (Link or DOI) β€”
Concentration Ranges
  Ace 0 – 0.5
  Ala 0.1 – 1.6
  Asc 0.4 – 1.7
  Asp 1 – 2
  Cr 4.5 – 10.5
  GABA 1 – 2
  GPC 0.4 – 1.7
  GSH 1.7 – 3
  Glc 1 – 2
  Gln 3 – 6
  Glu 6 – 12.5
  Gly 0.2 – 1
  Ins 4 – 9
  Lac 0.2 – 1
  Mac 0 – 10
  NAA 7.5 – 12
  NAAG 0.5 – 2.5
  PCho 0.2 – 1
  PCr 3 – 5.5
  PE 1 – 2
  Tau 3 – 6
  sIns 0.2 – 0.5
T2 Relaxation Ranges not per metabolite: one Lorentzian decay for all basis functions, exp(-gamma*t) with gamma in [2, 15] 1/s (1/gamma 66.67-500 ms, FWHM gamma/pi 0.6366-4.775 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 β€”
  Diffusion Model (Name, DOI) β€”
  fMRS Target Metabolites and Percent Changes β€”
Macromolecule
  Available in Isolation yes
  Simulated, Modeled, Experimental, or Mixture Modeled from experimental data: a fit of the cohort-average metabolite-nulled spectrum, 2DJ PRESS (TE 25, 30 and 35 ms, inversion time 900 ms) in occipital grey matter (MarjaΕ„ska et al., Magn Reson Med 2021, doi:10.1002/mrm.28942); a basis function sampled like the metabolites
  Description Mac
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 0.60-4.45 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 -60 – 20
    Im 1 -80 – 30
    Re x -100 – 60
    Im x -60 – 100
    Re x^2 -160 – 20
    Im x^2 -40 – 100
Noise
  Available in Isolation no
  Distribution additive complex Gaussian in the frequency domain (on the unnormalised FFT spectrum), real and imaginary parts independent; returned only summed with the random walk and random peaks
  Amplitude Scaling
    std of each of the real and imaginary parts, a.u., fixed 353.553
B0 Map
  B0 Map Model Used β€”
  Available in Isolation β€”
  Applied no
Tissue Segmentation
  Brain Map Model Used β€”
  Voxel Percentage(s) β€”
  Applied no
Artifacts
Zero-Order Phase
  Phase Range
    rad, as exp(-1j*phi0) -0.2 – 0.2
    deg -11.46 – 11.46
  Applied yes
First-Order Phase
  Phase Range
    rad/Hz, as exp(-1jphi1f) -1e-05 – 1e-05
    deg/ppm -0.0706 – 0.0706
  Amplitude Range β€”
  PPM Reference [ppm] -11.58
  Applied yes
Frequency Shifts - Global
  Frequency Shift Range
    rad/s, as exp(-1jepst) -5 – 5
    Hz -0.7958 – 0.7958
    ppm -0.006458 – 0.006458
  Applied yes
Residual Water
  Model Used (Equation, Description, or DOI) β€”
  If Analytical, Parameter Ranges β€”
  Available in Isolation β€”
  Applied no
  Chemical Shift [ppm] β€”
  Number of Subcomponents β€”
  Amplitude Scaling [a.u.] β€”
Unsuppressed Water
  Model Used (Equation, Description, or DOI) β€”
  Available in Isolation β€”
  Applied no
Lipid
  Model Used (Equation, Description, or DOI) β€”
  Available in Isolation β€”
  Applied no
Eddy Currents
  Model Used (Equation, Description, or DOI) β€”
  If Analytical, Parameter Ranges β€”
  Available in Isolation β€”
  Applied no
Out-of-Voxel
  Model Used (Equation, Description, or DOI) β€”
  If Analytical, Parameter Ranges β€”
  Available in Isolation β€”
  Applied no
Motion
  Model Used (Equation, Description, or DOI) β€”
  If Analytical, Parameter Ranges β€”
  Available in Isolation β€”
  Applied no
Random Walk (added)
  Model Used (Equation, Description, or DOI) smoothed random walk added to the whole frequency-domain spectrum, real and imaginary parts drawn independently (utils/components.py randomWalk): starts uniform in [low, high], Gaussian steps of std 'scale', a step that would leave [low, high] is reversed, then a moving average over 'smooth' points
  If Analytical, Parameter Ranges
    scale (step std), a.u. 0 – 1000
    smooth (moving-average window), points 1 – 1000
    low limit, a.u. -10000 – 0
    high limit, a.u. 0 – 10000
  Available in Isolation no
  Applied yes
Random Peaks (added)
  Model Used (Equation, Description, or DOI) Gaussian peaks added to the whole frequency-domain spectrum at uniformly random points (utils/components.py randomPeak): real part a Gaussian, imaginary part its Hilbert transform, times exp(-1j*phase)
  If Analytical, Parameter Ranges
    count per spectrum 0 – 4
    amplitude, a.u. 0 – 60000
    width (Gaussian std), points 0 – 100
    width (Gaussian std), Hz 0 – 195.3
    phase, rad 0 – 6.28319
  Available in Isolation no
  Applied yes
Data Quality
Signal Noise Ratio Range [a.u.]
  Range 16.1 – 44.3
  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 0.6366-4.775 Hz (Lorentzian) and 2*sqrt(ln2)*sigma/pi 1.06-7.95 Hz (Gaussian)
  Reference Metabolite and Moiety NAA, CH3 singlet (2.01 ppm)
  Reference Metabolite Linewidth Range [Hz] 3.04 – 12.34

Configuration 

{
  "acquisition": {
    "bandwidth": 4000.0,
    "centre": 4.65,
    "frequency": 123.22,
    "points": 2048
  },
  "baseline": true,
  "batch_statistics": true,
  "first": 1792,
  "last": 2036,
  "metabolites": [
    "Ace",
    "Ala",
    "Asc",
    "Asp",
    "Cr",
    "GABA",
    "GPC",
    "GSH",
    "Glc",
    "Gln",
    "Glu",
    "Gly",
    "Ins",
    "Lac",
    "Mac",
    "NAA",
    "NAAG",
    "PCho",
    "PCr",
    "PE",
    "Tau",
    "sIns"
  ],
  "model": "mrsuite.models.wand.WAND",
  "nuisance": [],
  "weights": null,
  "zero_fill": 1
}
Downloads last month
227
Safetensors
Model size
14.8M params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support