Pith. sign in

REVIEW 4 major objections 5 minor 10 references

MWF-MIMOSA for efficient simultaneous relaxometry and myelin water fraction mapping

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single 5-minute MRI scan can now produce whole-brain myelin water and relaxation maps.

desk verdict A well-engineered, honest methods paper that plausibly delivers fast simultaneous multi-parametric mapping, but the MWF values are not yet independently validated and should be treated as provisional. read the letter →

arxiv 2607.28984 v1 pith:CIAS2JBI submitted 2026-07-31 physics.med-ph

Yuting Chen , Yohan Jun , Hyeong-Geol Shin , Shizhuo Li , Shohei Fujita , Xingwang Yong , Jiye Kim , Jongho Lee
show 10 more authors
Gian Franco Piredda Tom Hilbert Aneri Bhatt Susie Y. Huang Huafeng Liu Huihui Ye Shahin Nasr Borjan Gagoski Kwok-Shing Chan Berkin Bilgic
This is my paper
classification physics.med-ph
keywords myelinwaterfractionquantitativeMRImulti-parametricmappingrelaxometryBlochsimulationself-supervisedreconstructionmagneticsusceptibilityseparationmultiplesclerosis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MWF-MIMOSA aims to show that myelin water fraction — the proportion of water trapped between myelin layers, a proxy for myelin content — can be mapped together with T1, T2, T2*, and magnetic susceptibility parameters from one acquisition, without the long scans and restrictive fixed-parameter assumptions of existing myelin water imaging. The authors extend the MIMOSA multi-parametric sequence with a short-T2-preparation block and a high-efficiency multi-echo gradient echo readout, then fit a three-compartment (myelin water, intra-/extra-axonal water, free water) Bloch-simulation model that estimates the myelin compartment's relaxation parameters rather than assuming them. A neural-network surrogate replaces the Bloch simulation in the fitting loop, giving a >100-fold speed-up in myelin water fraction estimation. In simulations the joint model beats both fixed-parameter and MGRE-only models in accuracy and precision, and in vivo it produces whole-brain 1-mm isotropic maps in about 5 minutes, with correlation to established myelin water imaging methods. If it holds up, this makes quantitative myelin imaging practical for clinical and high-resolution studies.

What carries the argument

The load-bearing object is the three-compartment signal model: the measured complex signal is a sum of Bloch-simulated signals from myelin water, intra-/extra-axonal water, and free water, each with its own fraction, T1, T2, T2*, and frequency shift (free water fixed at standard values). This model turns MWF estimation into a constrained joint optimization — fractions non-negative and sum to one, with total-variation smoothing on the myelin fraction only — rather than a fit with pre-fixed myelin relaxation parameters. To make the joint fit tractable, a multilayer perceptron with a weighted relative complex loss is trained to reproduce the Bloch simulation over physiologically relevant parame

What would settle it

A test that would settle the central claim: scan tissue with known myelin content (e.g., ex vivo samples or an animal model) where magnetization transfer is varied or blocked, or where myelin-water relaxation parameters lie outside the trained ranges (T1,MW 100–400 ms, R2,MW 25–100 Hz, R2*,MW 50–200 Hz); if MWF-MIMOSA estimates shift systematically with MT or with out-of-range relaxation while true myelin content is fixed, the no-MT and finite-range assumptions are violated. Comparing MWF-MIMOSA with histology-based myelination across regions with different MT would reveal such bias.

Watch

Extended reading notes

Core claim

The paper's central claim is that adding a short-TE T2-preparation module and a bipolar multi-echo gradient-echo readout to the MIMOSA sequence yields enough contrast to separate three water compartments — myelin water, intra-/extra-axonal water, and free water — and to estimate the myelin water fraction by fitting a Bloch-simulation signal model in which the myelin compartment's T1, T2, T2*, and frequency shift are free parameters rather than fixed literature values. This joint-estimation strategy is argued to be better conditioned than either fixed-parameter multi-contrast fitting or MGRE-only fitting, and to remain accurate across fiber orientations. The authors further claim that replaci

Load-bearing premise

The model assumes no magnetization transfer between water pools and fixes free-water relaxation values (T1=4500 ms, T2=500 ms, T2*=500 ms); if those assumptions fail — as they may in edema, disease, or dense white matter — the myelin water fraction estimates will be biased.

Editorial extensions

If this is right

  • If correct, a single ~5-minute scan can replace separate acquisitions for relaxometry, QSM, susceptibility-source separation, and myelin water imaging, reducing exam time and motion mismatch.
  • MWF mapping at 1 mm isotropic (and 0.7 mm with 10 min) becomes feasible, opening cortical myeloarchitecture and small-structure studies.
  • Because myelin-compartment relaxation parameters are estimated rather than assumed, the method may be less biased in pathology where tissue relaxation changes.
  • The reported strong correlation with MCR-MWI (r=0.924) and moderate correlation with GRASE (r=0.516) suggest the maps measure overlapping but not identical tissue properties, meaning MWF-MIMOSA could serve as a higher-resolution complement.
  • The single-subject MS demonstration suggests a clinical route to simultaneously assessing demyelination (low MWF), edema (high IEW/FW), and prolonged relaxation in lesions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that fixing free-water relaxation values makes MWF vulnerable when free-water properties deviate; a natural test is re-fitting with free-water relaxation free or constrained by measurement.
  • Because magnetization exchange between water pools is not modeled, MWF likely absorbs some magnetization-transfer weighting; a Bloch-McConnell extension would reveal the sign and magnitude of this bias.
  • The MLP surrogate's finite training ranges imply that out-of-range disease states could bias estimates; a practical safeguard is to detect out-of-range inputs and fall back to Bloch simulation.
  • The cortical MWF dynamic range of 1–3% suggests that high-resolution cortical myelin mapping may need dedicated noise and regularization tuning before group comparisons.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes MWF-MIMOSA, an extension of the MIMOSA quantitative MRI framework that adds a short-TE T2-preparation module and a bipolar multi-echo gradient-echo readout, enabling simultaneous T1, T2, T2*, QSM, susceptibility source separation, and myelin water fraction (MWF) mapping from a single acquisition. Parameter estimation uses a three-compartment Bloch-simulation model (myelin water, intra-/extra-axonal water, free water) with the compartments' relaxation parameters jointly estimated, and a multilayer perceptron surrogate trained within the GACELLE framework to accelerate the Bloch simulations by more than 100-fold. The authors report numerical simulations comparing the proposed joint-estimation model against a preparation-only model and an MGRE-only model, and in-vivo acquisitions in five healthy subjects and one MS patient, showing strong correlation with MCR-MWI (r=0.924) and moderate correlation with GRASE (r=0.516). Whole-brain maps are demonstrated at 1 mm isotropic in about 5 minutes and at 0.7 mm isotropic in about 10 minutes. The paper includes open source code, Pulseq sequence files, and indicates a representative raw dataset is publicly available.

Significance. If the accuracy claims hold, this would be a practically valuable contribution: a single, short, high-resolution acquisition yielding multiple quantitative contrasts plus MWF, with reconstruction and estimation pipelines that are reproducible and openly documented. The engineering is substantial: the sequence modification, the self-supervised reconstruction, the MLP surrogate with carefully ablated loss design, and the GPU-accelerated fitting together form a coherent and nontrivial system. The numerical experiments are internally consistent, and the in-vivo comparisons show expected spatial patterns. However, the biophysical specificity and absolute accuracy of the MWF estimates rest on modeling assumptions — notably the omission of magnetization exchange and the finite, in-distribution training of the MLP — that are acknowledged but not quantitatively bounded. The paper's own Discussion states that magnetization exchange was not modeled and may bias myelin-related parameter estimates, and the MLP validation uses synthetic data sampled from the same parameter ranges used for training. Because the correlations with MCR and GRASE could be dominated by shared spatial priors, they do n

major comments (4)
  1. [Section 4 / Eq. (2)] The signal model in Eq. (2) omits magnetization exchange between water pools, and the Discussion explicitly states that this 'may have introduced bias into the estimation of myelin-related parameters.' This is load-bearing because MWF is estimated as the amplitude of the short-T2* component; unmodeled MT can change the apparent weight of that component. The in-vivo correlations with MCR (r=0.924) and GRASE (r=0.516) cannot detect a bias shared across methods. Please quantify the sensitivity of MWF-MIMOSA to MT, e.g., by simulating data with a Bloch-McConnell model and fitting with the current model, or by adding MT as a nuisance parameter in a subset of simulations.
  2. [Supplementary §1.1 and §1.3 / Eq. (3)] The MLP surrogate is trained and validated only on parameter ranges given in Supp. §1.1 (e.g., T1,MW 100–400 ms, R2,MW 25–100 Hz, R2*,MW 50–200 Hz), and §1.3 states validation used the same ranges. Eq. (3) constrains only compartment fractions, not relaxation parameters, so the optimizer can drive inputs outside the surrogate's training support. This matters precisely for the MS application, where relaxation times are expected to shift. Please add explicit bounds or penalty terms on all fitted parameters, and report out-of-distribution performance of the MLP (e.g., NRMSE on parameter ranges extended beyond the training ranges) to quantify extrapolation error.
  3. [Section 2.5] The numerical simulations synthesize data with the same Bloch-based three-compartment model that is then used for fitting. This is a self-consistency and conditioning check, not an independent test of absolute accuracy against more complete biophysical physics. The simulation results therefore cannot address model-misspecification errors such as MT exchange, fixed free-water relaxation, or incorrect compartmental line shapes. Please either rephrase the claims as precision/conditioning comparisons or add simulations with an independent generative model (e.g., Bloch-McConnell or an extended multi-pool model) to assess bias under model mismatch.
  4. [Section 2.4.1] The free-water compartment is fixed to T1,FW=4500 ms, T2,FW=500 ms, T2*,FW=500 ms, and Δω_FW=0. These values are reasonable for pure CSF but may be misspecified in voxels with partial-volume or pathological water content (e.g., edema, lesion). Since the fractions must sum to unity, misspecification in the FW compartment can directly shift the estimated MWF. Please include a sensitivity analysis varying the FW relaxation values (or, alternatively, fitting them with weak priors) to show the impact on MWF in healthy white matter and in the MS lesion.
minor comments (5)
  1. [Figure 2(a)] Typo: 'Phase images are used further used for quantitative susceptibility mapping.' Please delete the duplicated 'used'.
  2. [Supplementary §1.1] The parameter list includes '△ω_IEW ∈ [-0.1, 0.05] ppm' twice. Please remove the duplicate.
  3. [Figure 1(a) vs Section 2.6.2] Figure 1 reports total sequence TR of 6,956.5 ms, while the in-vivo protocol in Section 2.6.2 lists total sequence TR = 6,055 ms. Please clarify which TR applies to the in-vivo acquisitions and whether the Bloch simulations used the 6,956.5 ms value.
  4. [Section 2.6.4 / Figure 7] The correlation analysis uses 17 ROI-mean values but aggregates across five subjects, so the effective sample size is smaller than 17 and the p-values may be optimistic. Consider reporting subject-level correlations or a mixed-effects model that accounts for repeated measures.
  5. [Section 4] The discussed limitations — MT exchange not modeled, limited literature on compartmental relaxation parameters, and a single MS patient — are appropriately acknowledged, but they should be connected more explicitly to the validation claims: the Discussion currently states that MT 'may have introduced bias' after the Results section has already presented MWF as an accurate and robust measure. A short paragraph stating the conditions under which the MWF estimates are expected to be unbiased would improve interpretability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central MWF derivation is a forward-model fit validated against external benchmarks; self-citations are tool citations, not load-bearing reductions.

full rationale

The paper's derivation chain is: acquire multi-contrast MIMOSA data, reconstruct with MZS-SSL, estimate single-compartment maps by dictionary matching, then estimate MWF by fitting the three-compartment forward model of Eq. (2) via Eq. (3). MWF is the fitted fraction f_MW; it is not defined in terms of the output or fitted to the validation data. The MLP is a computational surrogate for Bloch simulations, trained on Bloch-simulated labels and used only to speed up the fitting; its reported 0.17% NRMSE on an in-distribution test set is an interpolation check of the surrogate, not evidence of biophysical validity. The paper does not rest its biophysical accuracy claim on that NRMSE: it separately provides ground-truth simulations with a digital phantom and in-vivo comparisons against MCR and GRASE, which are external to the fitted MWF values. The acknowledged omission of magnetization exchange and fixed free-water relaxation values are modeling limitations that could bias MWF estimates in some tissues, but they are explicitly stated as limitations rather than hidden assumptions, and they do not make the derivation self-referential. The several self-citations (MIMOSA, GACELLE, MZS-SSL, chi-sepnet) are references to prior tools and sequence designs, not to a uniqueness theorem or to an unverified premise that forces the present result. No equation or fitted parameter is repackaged as an independent prediction. Therefore, no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central MWF estimate rests on compartment fractions and relaxation/frequency parameters fitted under a three-compartment Bloch model with fixed free-water values, plus an empirically chosen TV weight. No new physical entities are introduced; the MLP surrogate is a computational approximation whose training distribution is a load-bearing assumption.

free parameters (5)
  • Compartment fractions f_MW, f_IEW, f_FW = per-voxel maps (not scalar)
    Central MWF output; estimated by optimization in Eq. 3, with f_FW = 1 - f_MW - f_IEW.
  • Compartmental relaxation parameters T1, R2, R2* for MW and IEW = per-voxel maps; ranges in MLP training (T1,MW 100-400 ms, R2,MW 25-100 Hz, R2*,MW 50-200 Hz, etc.)
    Jointly estimated to avoid fixing MW relaxation values; a load-bearing part of the three-compartment model.
  • Compartmental frequency shifts Δω_MW, Δω_IEW = per-voxel maps; training ranges Δω_MW ∈ [-0.05, 0.25] ppm, Δω_IEW ∈ [-0.1, 0.05] ppm
    Estimated from MGRE phase information; included in the signal model (Eq. 2).
  • TV regularization weight λ = 3 × 10^-4
    Selected based on pilot experiments; fixed for all datasets in Section 2.4.1.
  • WRCL loss hyperparameters τ and ε = τ = 10^-2, ε = 10^-4
    Chosen via ablation study in Supplementary Section 1.4; affect MLP surrogate accuracy.
assumptions (5)
  • domain assumption Bloch simulations accurately describe the MRI signal evolution for the MWF-MIMOSA sequence.
    Eq. (2) defines the forward model through a Bloch simulation operator for the sequence timing.
  • domain assumption The brain tissue signal is composed of three non-exchanging water compartments (MW, IEW, FW).
    Section 2.4.1 introduces the three-compartment model; magnetization exchange is explicitly not modeled in the Discussion.
  • domain assumption Free-water relaxation values are fixed constants (T1,FW = 4500 ms, T2,FW = 500 ms, T2*,FW = 500 ms) and Δω_FW = 0.
    Section 2.4.1 states these fixed values, citing prior MRF/partial-volume literature.
  • domain assumption The hollow cylinder fiber model (HCM) adequately describes orientation-dependent compartmental frequency shifts in white matter.
    Section 2.5 uses HCM to simulate frequency shifts at 0°, 45°, and 90° fiber orientations.
  • ad hoc to paper The MLP training parameter ranges cover the relevant in-vivo tissue space.
    Supplementary Section 1.1 defines finite ranges for T1, R2, R2*, and frequency shifts; values outside these ranges are extrapolation and not validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MWF-MIMOSA for efficient simultaneous relaxometry and myelin water fraction mapping." pith.science (2026). https://pith.science/paper/CIAS2JBI

@misc{pith2026260728984,
  author       = {Pith},
  title        = {Pith review of: MWF-MIMOSA for efficient simultaneous relaxometry and myelin water fraction mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIAS2JBI}},
  note         = {Machine review of arXiv:2607.28984}
}
read the original abstract

Quantitative magnetic resonance imaging (qMRI) provides improved sensitivity and specificity to tissue composition and pathological alterations compared with conventional contrast-weighted imaging. Among various qMRI biomarkers, myelin water imaging is of particular interest because myelin plays a central role in brain function and its alteration is closely associated with many neurological diseases. However, conventional myelin water fraction (MWF) mapping techniques are often limited by long scan times, low spatial resolution, reduced signal-to-noise ratio (SNR), and high specific absorption rate (SAR). Here, we propose MWF-MIMOSA for efficient simultaneous T1, T2, T2* mapping, magnetic susceptibility source separation, and MWF estimation. To achieve this, multi-contrast and multi-slice zero-shot self-supervised learning (MZS-SSL) was used to jointly reconstruct whole-brain complex-valued images. To improve computational efficiency of the parameter estimation step, a multilayer perceptron (MLP) was trained within the GACELLE GPU-accelerated parameter estimation framework to circumvent the computationally intensive Bloch simulation process, resulting in a >100-fold computational speed-up in MWF estimation. Numerical simulations were performed to evaluate the accuracy and precision of MWF-MIMOSA, and in-vivo results further demonstrated its robustness. Comparison with existing myelin water imaging methods showed that MWF-MIMOSA is highly correlated with established approaches, while providing complementary quantitative parameter maps at higher spatial resolution and with shorter scan times. Notably, simultaneous multi-parametric mapping was achieved in 5 min at 1 mm isotropic resolution, and in 10 min at 0.7 mm isotropic resolution. These results demonstrate the potential of MWF-MIMOSA for fast, high-resolution simultaneous relaxometry and myelin water imaging.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 1 canonical work pages

  1. [1]

    Avants, B.B. et al. (2008) ‘Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain’, Medical image analysis , 12(1), pp. 26–41. Avants, B.B. et al. (2011) ‘A reproducible evaluation of ANTs similarity metric performance in brain image registration’, NeuroImage , 54(3), pp. 20...

    arXiv 2008

  2. [9]

    Available at: https://doi.org/10.7554/eLife.61523. Ma, S. et al. (2025) ‘Advances in neuroimaging applications of quantitative susceptibility mapping’, Meta-Radiology , 3(3), p. 100148. Mohammadi, S. and Callaghan, M.F. (2021) ‘Towards in vivo g-ratio mapping using MRI: Unifying myelin and diffusion imaging’, Journal of neuroscience methods , 348, p. 1089...

    work page doi:10.7554/elife.61523 2025

  3. [10]

    For each mini-batch voxel, f MW and f IEW were randomly sampled under the constraint f MW + f IEW ≤ 1 and the f FW was set as 1- f MW - f IEW

    MLP-Based Bloch Surrogate Model for MWF Estimation 1.1 Training data Similar to the EPG-X approximation model implemented in GACELLE (Chan et al., 2025), around 10 7 random parameter sets were generated during training with the following parameter ranges: f MW ∊ [0, 0.3], f IEW ∊ [0, 1], T 1,MW ∊ [100, 400] ms, T 1,IEW ∊ [700, 1500] ms, R 2,MW ∊ [25, 100]...

    2025

  4. [102]

    and Pike, G.B

    Kwan, R.K., Evans, A.C. and Pike, G.B. (1999) ‘MRI simulation-based evaluation of image-processing and classification methods’, IEEE transactions on medical imaging , 18(11), pp. 1085–1097. Lancione, M. et al. (2024) ‘Magnetic resonance fingerprinting-based myelin water fraction mapping for the assessment of white matter maturation and integrity in typica...

    1999

  5. [136]

    Feng, J. et al. (2025) ‘Spatiotemporal Implicit Neural Representation for Unsupervised Dynamic MRI Reconstruction’, IEEE transactions on medical imaging , 44(5), pp. 2143–2156. Feng, R. et al. (2026) ‘Accelerating Multiparametric Quantitative MRI Using Self-Supervised Scan-Specific Implicit Neural Representation With Model Reinforcement’, Magnetic resonan...

    2025

  6. [914]

    Kim, M. et al. (2025) ‘χ- sepnet: Deep Neural Network for Magnetic Susceptibility Source Separation’, Human brain mapping , 46(2), p. e70136. Kim, W. et al. (2023) ‘χ- Separation Imaging for Diagnosis of Multiple Sclerosis versus Neuromyelitis Optica Spectrum Disorder’, Radiology , 307(1), p. e220941. Klein, A. et al. (2009) ‘Evaluation of 14 nonlinear de...

    2025

  7. [1991]

    6254–6269

    , 30(12), pp. 6254–6269. Faizy, T.D. et al. (2020) ‘The Myelin Water Fraction Serves as a Marker for Age-Related Myelin Alterations in the Cerebral White Matter - A Multiparametric MRI Aging Study’, Frontiers in neuroscience , 14, p

    2020

  8. [1997]

    , 329, p. 107011. Graham, S.J. and Henkelman, R.M. (1997) ‘Understanding pulsed magnetization transfer’, Journal of magnetic resonance imaging : JMRI , 7(5), pp. 903–912. Griswold, M.A. et al. (2002) ‘Generalized autocalibrating partially parallel acquisitions (GRAPPA)’, Magnetic resonance in medicine , 47(6), pp. 1202–1210. Heckel, R. et al. (2024) ‘Deep...

    1997

Show all 10 references
  1. [5467]

    (2012) ‘FreeSurfer’, NeuroImage , 62(2), pp

    Fischl, B. (2012) ‘FreeSurfer’, NeuroImage , 62(2), pp. 774–781. Fonov, V. et al. (2011) ‘Unbiased average age-appropriate atlases for pediatric studies’, NeuroImage , 54(1), pp. 313–327. Frost, R. et al. (2016) ‘Prospective motion correction and selective reacquisition using ...

    2012

  2. [7228]

    Liao, C. et al. (2023) ‘High-resolution myelin-water fraction and quantitative relaxation mapping using 3D ViSTa-MR fingerprinting’, ArXiv . Available at: https://www.ncbi.nlm.nih.gov/pubmed/38196746. Li, T.-Q. et al. (2009) ‘Characterization of T(2)* heterogeneity in human br...

    2023

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.