Skip to the content
BlochTwinRequest access

Pulseq in.
Realistic raw data out.

BlochTwin is an MRI simulator for research. It runs your pulse sequence on a head phantom built from real human data and returns what a scanner would have recorded: multi-coil k-space with correlated receiver noise, the reconstructed images, and the ground truth behind them.

Free for academic research in 2D. Access opens to invited groups first.

We keep your email address, its domain and the affiliation you choose only to run the access list and to write to you about access. We delete them when you ask (contact@newmean.ai); 12 months after we tell you your access is open, if you have not signed up; 24 months after you asked, if we have not invited you; and after 7 days if you do not confirm the address. You can use BlochTwin’s public pages without giving them. Privacy policy

Working at a company? Ask about an evaluation.

Simulated T2-weighted turbo spin echo image of an axial slice of the head at the level of the lateral ventricles, with visible receiver noise
3 T32 channelsT2 TSETR 4000 ms, TE 96 ms1 × 1 × 4 mm, 0:44
Raw k-space of the same scan: a bright centre with streaks, logarithmic magnitude summed over the 32 channels
32 × 440 × 220samplesraw k-spacelog magnitude, all channels
The 32 images seen by the individual receive channels, each bright near its own coil
the 32 channels, each on its own
The same image simulated without receiver noise
receiver noise off
The same scan with the head nodding by one degree: ghosting along the phase-encoding direction
1° nod during the scan
Figure 1. One simulated scan and what comes back with it. (a) A T2-weighted turbo spin echo of the head phantom at 3 T. The slice profiles of its own RF pulses, a B1+ map, a 32-channel receive array and correlated receiver noise are in the signal; none is added to a finished image. (b) Its raw k-space as the receiver recorded it. (c) What each of the 32 channels sees. (d) The same scan with the noise switched off, a reference no scanner can give you. (e) The same scan while the head nods by 1°.

From a sequence file to raw data

  1. 1.Give it a sequence

    Upload a Pulseq file. Every RF pulse, gradient and ADC event is simulated as written, so a new sequence needs no new code on our side.

    Or start from a built-in gradient echo, spin echo, turbo spin echo, FLAIR or EPI and set its parameters.

    RFGzGyGxADC010203040 ms
    Figure 2a. The sequence behind Figure 1 as the engine read it from the Pulseq file: the excitation, the first five of the 20 refocusing pulses and four readouts, the first 46 ms of a 4 s repetition. The whole file has 1,804 blocks.
  2. 2.Choose the head and the scanner

    The phantom stores how much of each tissue every voxel holds, so a voxel on a boundary is a mixture, as in a real head. At the field strength you choose (1.5, 3 or 7 T) each tissue takes its T1, T2 and T2* from literature tables. At 3 T, the default, the deep nuclei’s T1 comes from measured T1 maps of 16 adults.

    The susceptibility field, the transmit field and the receive array are computed for that same head.

    Map of T1 over the slice: white matter shortest, grey matter longer, cerebrospinal fluid longest
    0.32.6 sT1
    Map of the static field offset after shimming: near zero inside the brain, strong behind the frontal sinus and along the scalp
    −60+60 HzΔB0 after shim
    Map of the transmit field: highest in the middle of the head, lower towards the surface
    0.801.25B1+
    receive sensitivity4 of 32 loops
    Figure 2b. What the simulation is given for the slice of Figure 1, at 3 T: the phantom’s T1, the field offset left after second-order shimming (the dipole pattern in front comes from the air of the frontal sinus), the transmit field, and the sensitivity of four of the receive loops.
  3. 3.Download what the receiver recorded, and what it could not

    The raw data arrive in the format your reconstruction already reads. Beside them comes everything the simulation knew and a scanner never does: the answer your method is trying to recover.

    You getFormat
    As from a scanner
    Raw k-space, every channel, with a noise scanISMRMRD
    Reconstructed imagesNIfTI, PNG
    Only from a simulation
    The same raw data without noiseISMRMRD
    The receive sensitivities the simulation usedNIfTI
    Tissue fractions and parameter mapsNIfTI
    The head pose at every k-space line, if it movedJSON
    The predicted sound of the gradientsWAV
    Run record: engine version, seed, every parameterJSON

The same three steps from Python, for a parameter sweep or a notebook. Submitting and viewing in the browser work without it.

import blochtwin as bt

run = bt.simulate(
    sequence="my_tse.seq",    # any 2D Pulseq file
    phantom="head-1mm",
    field_strength=3.0,       # tesla
    coils=32,
    noise=True,
    seed=7,
)
run.download("raw.h5")        # ISMRMRD, 32 channels, noise scan included
truth = run.noise_free_kspace()
Listing 1. The Python client is in preparation; names may change before it ships.

Look at the output before you sign up

Seven scans of one slice, computed by the engine and shown as they came out. Pick one and read its k-space next to it.

T2 turbo spin echo: reconstructed image of an axial slice through the lateral ventricles
image
T2 turbo spin echo: raw k-space, logarithmic magnitude over all channels
k-space, log magnitude
Contrast
The T2 scan with one thing changed

Cerebrospinal fluid bright, the echo train's T2 decay in k-space.

TR / TE
4000 / 96 ms
Echo train
20 echoes, 9.6 ms apart
Matrix
220 × 220 over 220 mm
Scan time
0:44
Field
3 T
Receive
32 channels
Slice
axial, 4 mm
Figure 3. Precomputed results on the head phantom at 3 T with the 32-channel array. The images are the root-sum-of-squares over the channels, divided by the array’s own sensitivity as a prescan normalisation does. A display filter (Hamming apodisation, twofold interpolation) is on except for the motion pair.

What is in the signal

Each effect enters where it does in a scanner: in the magnetisation, in the field or in the receive chain. All of them are available on every plan.

Magnetisation

  • Bloch simulation of any Pulseq sequence: RF pulses as rotations in 10 µs steps, exact precession and relaxation in between (Bloch 1946; Pauly 1991; Layton 2017).
  • Slice profiles that follow from the RF waveform and its gradient, and cross-talk between neighbouring slices.
  • Reversible dephasing inside a voxel (T2′): gradient echoes decay with T2*, spin echoes refocus it (Endres 2024).
  • Tissue values at 1.5, 3 and 7 T from literature tables (Rooney 2007); at 3 T the deep nuclei’s T1 from MP2RAGE maps of 16 adults, and T2* of white matter, cortex and deep nuclei from a 12-echo gradient echo of 18 adults (OpenNeuro ds004958). The variation inside each tissue is taken from the person the phantom was built from.
  • Fat with a six-peak spectrum, so chemical shift and its dephasing appear on their own (Yu 2008).

Fields and the subject

  • ΔB0 from the head’s own susceptibility, air cavities included, after centre-frequency and second-order shim (Salomir 2003; Marques 2005).
  • A transmit-field map with the central brightening of a head at 3 T (a smooth model, not a field calculation).
  • Rigid head motion during the scan, line by line in k-space (Zaitsev 2015).
  • Breathing as a field fluctuation; blood that pulses with the heartbeat, with its inflow enhancement and wash-out.
  • The sound of the sequence, from its gradient waveforms through a transfer model built on published measurements of gradient coils (Hedeen 1997).

Receive chain

  • Array sensitivities from the fields of the loops, by reciprocity; 1, 8 or 32 channels on a helmet (Hoult 1976).
  • Thermal noise added to k-space per channel, correlated between channels through the conducting head (Roemer 1990). Rician statistics in magnitude images follow from it (Gudbjartsson 1995).
  • A noise scan in front of the data, as a scanner records it.
  • Reconstruction you can replace: root-sum-of-squares with prescan normalisation or noise prewhitening, SENSE (Pruessmann 1999), partial Fourier.

Not modelled yet. Gradient delays and eddy currents, concomitant fields, coupling between receive coils, RF wavelength effects at 7 T, diffusion and magnetisation transfer. Head motion, breathing and pulsing blood work with the built-in sequences today and not yet with an uploaded Pulseq file. A whole-head 3D scan takes hours to simulate at 2 mm and far longer below it, which is why the free plan starts with 2D.

Checked against closed forms, another simulator and a scanner

The source is closed, so the evidence has to be open. The band for a pass was fixed before the measured data were looked at, and what falls outside it stays in the table.

ComparedAgainstResultVerdict
Spin echo, inversion recovery and flip-angle signals of single voxelsClosed-form solutions of the Bloch equationsIdentical to machine precisionagrees
The same experiments, white matter, grey matter and CSFKomaMRI (Castillo-Passi 2023)Mean absolute difference 0.0023 %, worst 0.0113 %agrees
Dephasing inside a voxel across a spin-echo readoutexp(−t/T2) exp(−R2′|t − TE|)Within 0.2 %, cusp includedagrees
Receiver noise after prewhiteningThe covariance that was put inChannel correlation below 0.02; covariance recovered within 2 %agrees
Noise correlation between the channels of a 32-channel head arrayNoise scan of a 7 T head coil (Lüsebrink 2021)Measured mean |ρ| 0.06; model 0.12 to 0.25too high, 2 to 4 times
MPRAGE at 7 T, same person: white-matter signal, 5° scan over 1° scanThe scanner’s imagesMeasured 5.09, predicted 5.02agrees
Same scans: where the boundary between grey and white matter lies, and how sharp it isThe scanner’s imagesPosition within 0.02 mm, width 0.47 against 0.46 mmagrees
Same scans: contrast to white matter, predicted over measured. Pass band 0.9 to 1.1The scanner’s imagesCortical grey matter 0.93inside the band
Caudate 0.89just outside
Putamen 0.82, thalamus 0.75, pallidum 0.73, CSF 1.43outside the band
Table 1. Checks so far. The measured data are the 7 T scans of one volunteer published by Lüsebrink and colleagues (OpenNeuro ds003563). The phantom was built from other sessions of the same person. Tissue values come from literature tables. Only the geometry was matched to the scans: a rigid transform, and a smooth warp for the scanner’s uncorrected gradient distortion, both estimated between two real images.

Where it does not match

Deep nuclei. Caudate, putamen, thalamus and pallidum are all brighter in the scan than predicted: in this person their T1 is 5 to 19 % shorter than the literature table. The phantom carries each tissue’s variation, not yet this person’s own tissue means.

Cortex. Measured grey matter is brighter next to white matter than at the outer surface (0.62 falling to 0.54 of white matter); the prediction is nearly flat at 0.52 to 0.54. The phantom has one T1 through the depth of the cortex.

Transmit field. The comparison assumes it uniform. At 7 T it is not, and these scans cannot tell us what it was.

Outside the brain. Scalp, face and neck were left out: the shading correction is fitted to white matter and does not hold there.

Plane
Noise in the simulated image
Simulated MPRAGE of the brain
Measured MPRAGE of the same brain at 7 T
measured7 T scanner imagesimulatednoise as measured
Figure 4. The same person, measured at 7 T and simulated (MPRAGE, 0.7 mm, brain only). Drag the divider. Both images are divided by their smooth white-matter shading, so white matter is 1 in each. The simulated side has no vessels and a flat cortex, which is where it differs most from the scan; the noise is the level measured from two repeats of the scan.

What it is for

Trying a sequence before the scanner is free

Contrast, slice profile, what off-resonance and motion do to it, and how loud it will be: from the file you would take to the scanner.

Testing reconstruction and learning methods

Multi-coil raw data with the noise-free k-space, the sensitivities and the tissue maps they came from. Change one factor at a time (noise, motion, field strength) and score against the truth.

Checking a quantitative method

The T1, T2 and proton density of every voxel are known, so a mapping sequence and its fit can be compared with the values that went in.

Free for academic research in 2D

The plans differ in dimensions, phantoms, compute and what you may do with the results. The physics is the same on all of them. The free plan opens first.

PropertyCommunityResearcher and LabIndustry
ForResearchers at universities, institutes and hospitalsThe same, when 2D is not enoughCompanies
PriceFreeTo be announcedBy contract, after an evaluation
Dimensions2D, multi-slice included2D and 3D2D and 3D
PhantomsShepp–Logan, vial phantom, procedural head, a real head at 1 mmAdds the 0.5 mm head, lesions and, for a lab, your own phantomsAdds phantoms built to order
Physics optionsAllAllAll
ComputeA monthly allowance; the cost of a run is shown before it startsA larger allowance, several runs at onceBy contract
Ways inBrowser, API, Python clientAdds batch submissionAdds dataset generation
Use of resultsNon-commercial research, with a citationThe sameCommercial use, training data included
OpensFirst, to invited groups, then to everyoneAfter the free planEvaluation by approval, seven days
Table 2. A sequence counts as 2D when its readouts fill two dimensions of k-space; a stack of 2D slices is still 2D.

Questions

The source is closed. Why should I trust the output?

Because you can check it without reading our code. Sequences go in as Pulseq files and data come out as ISMRMRD and NIfTI, so you can reconstruct the raw data with your own tools and set it beside your own scans. The physics is documented with its sources, and the table above lists every check, the failed ones included. Each result records the engine version, the random seed and every parameter.

What happens to a sequence I upload?

It stays yours. We use it to run your simulation and for nothing else, and we delete it when you ask. The terms of use will say so in full before sign-up opens.

How do I cite it?

Every result comes with a citation line and its run record. A technical report on the engine and its validation is in preparation; cite that once it is out.

Will the same run give the same numbers next year?

A result belongs to an engine version. Older versions stay available for re-running for a fixed period, which the terms of use will state.

Who counts as academic?

Anyone doing non-commercial research at a university, a research institute or a hospital who signs in with that institution’s email address. An address we cannot recognise automatically is looked at by a person.

Can a company use it?

Yes, on the Industry plan. An evaluation runs for seven days once it is approved. Results from an evaluation or from an academic plan may not be used commercially.

Is it a medical device?

No. BlochTwin is a research tool. Its output must not be used for diagnosis or for decisions about treatment.

Request access with your institutional email

Free for academic research in 2D. Access opens to invited groups first.

We keep your email address, its domain and the affiliation you choose only to run the access list and to write to you about access. We delete them when you ask (contact@newmean.ai); 12 months after we tell you your access is open, if you have not signed up; 24 months after you asked, if we have not invited you; and after 7 days if you do not confirm the address. You can use BlochTwin’s public pages without giving them. Privacy policy

For a company evaluation or anything else, write to contact@newmean.ai.

References

  1. Bloch F. Nuclear induction. Phys Rev 70, 460–474 (1946).
  2. Castillo-Passi C, et al. KomaMRI.jl: an open-source framework for general MRI simulations with GPU acceleration. Magn Reson Med 90, 329–342 (2023).
  3. Endres J, et al. Phase distribution graphs for fast, differentiable, and spatially encoded Bloch simulations of arbitrary MRI sequences. Magn Reson Med 92, 1189–1204 (2024).
  4. Gudbjartsson H, Patz S. The Rician distribution of noisy MRI data. Magn Reson Med 34, 910–914 (1995).
  5. Hedeen RA, Edelstein WA. Characterization and prediction of gradient acoustic noise in MR imagers. Magn Reson Med 37, 7–10 (1997).
  6. Hoult DI, Richards RE. The signal-to-noise ratio of the nuclear magnetic resonance experiment. J Magn Reson 24, 71–85 (1976).
  7. Layton KJ, et al. Pulseq: a rapid and hardware-independent pulse sequence prototyping framework. Magn Reson Med 77, 1544–1552 (2017).
  8. Lüsebrink F, et al. Comprehensive ultrahigh resolution whole brain in vivo MRI dataset as a human phantom. Sci Data 8, 138 (2021). OpenNeuro ds003563.
  9. Marques JP, Bowtell R. Application of a Fourier-based method for rapid calculation of field inhomogeneity due to spatial variation of magnetic susceptibility. Concepts Magn Reson B 25B, 65–78 (2005).
  10. Pauly J, et al. Parameter relations for the Shinnar–Le Roux selective excitation pulse design algorithm. IEEE Trans Med Imaging 10, 53–65 (1991).
  11. Pruessmann KP, et al. SENSE: sensitivity encoding for fast MRI. Magn Reson Med 42, 952–962 (1999).
  12. Roemer PB, et al. The NMR phased array. Magn Reson Med 16, 192–225 (1990).
  13. Rooney WD, et al. Magnetic field and tissue dependencies of human brain longitudinal 1H2O relaxation in vivo. Magn Reson Med 57, 308–318 (2007).
  14. Salomir R, et al. A fast calculation method for magnetic field inhomogeneity due to an arbitrary distribution of bulk susceptibility. Concepts Magn Reson B 19B, 26–34 (2003).
  15. Yu H, et al. Multiecho water-fat separation and simultaneous R2* estimation with multifrequency fat spectrum modeling. Magn Reson Med 60, 1122–1134 (2008).
  16. Zaitsev M, et al. Motion artifacts in MRI: a complex problem with many partial solutions. J Magn Reson Imaging 42, 887–901 (2015).