A Digital Phantom for MR Spectroscopy Data Simulation

A.T. Gudmundson; C.W. Davies-Jenkins; D.M.J. van de Sande; D. Simicic; G. Oeltzschner; G. Simegn; H.J. Z\"ollner; \.I. \"Ozdemir; J.P. Merkofer; R.A.E. Edden

REVIEW 3 major objections 6 minor 3 references

A Digital Phantom for MR Spectroscopy Data Simulation

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A modular digital brain phantom combines anatomical tissue maps with literature metabolite values to simulate MR spectroscopy spectra realistic enough to validate processing algorithms.

desk verdict Useful modular MRS simulator with a circular in-sample validation; the framework is a real contribution, but the realism claim is overstated. read the letter →

arxiv 2412.15869 v2 pith:HFCC5BWJ submitted 2024-12-20 physics.med-ph physics.bio-ph

D.M.J. van de Sande , A.T. Gudmundson , S. Murali-Manohar , C.W. Davies-Jenkins , D. Simicic , G. Simegn , İ. Özdemir , S. Amirrajab

show 4 more authors

J.P. Merkofer H.J. Zöllner G. Oeltzschner R.A.E. Edden

This is my paper · ORCID

classification physics.med-phphysics.bio-ph

keywords magneticresonancespectroscopydigitalbrainphantomspectralsimulationmetaboliteconcentrationsT2relaxationdataaugmentationsingle-voxel

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

The paper introduces a framework for generating synthetic magnetic resonance spectroscopy data: a digital brain phantom that layers tissue-specific metabolite concentrations and relaxation times onto anatomical brain templates, then feeds them through a signal model that adds metabolite, macromolecule, water, lipid, noise, and shim effects. The authors claim the resulting spectra closely match in-vivo spectra in shape, signal-to-noise ratio, and metabolite quantification, while also producing a wider range of variability useful for robustness testing and data augmentation. Because the outputs carry known ground-truth concentrations and are saved in a standard open format, the framework is positioned as a validation and training resource for MRS processing tools.

What carries the argument

The load-bearing mechanism is the MRS phantom: a three-dimensional map in which every voxel carries a tissue label (white matter, gray matter, CSF, or background) and each tissue label is associated with a metabolite dataframe of concentration means, standard deviations, and T2 values derived from a filtered meta-analysis. A signal model converts this phantom into spectra by summing tissue-weighted metabolite basis signals, a Voigt-line macromolecule background (a blend of Lorentzian and Gaussian line shapes), a five-component residual water signal, a triglyceride lipid signal weighted by a proximity-to-skull mask, complex Gaussian noise, and optional shim-induced line broadening. The modularity, with interchangeable anatomical templates, user-defined basis sets, and configurable parameters, is what lets the same phantom produce both in-vivo-like and deliberately degraded spectra.

What would settle it

Generate a simulated dataset with the pipeline's default or tuned parameters, then compare it against a separate in-vivo dataset acquired with the same sequence at a different site or subject group that was not used for any parameter optimization; if spectral shape, SNR, or quantified metabolite values diverge substantially from the in-vivo set, the realism claim would be refuted. A simpler check: run the simulated spectra through an independently developed quantification tool and see whether the ground-truth concentrations are recovered within expected error.

Watch

Extended reading notes

Core claim

The central discovery claimed is that realistic single-voxel MR spectroscopy data can be generated from a spatial phantom rather than from abstract spectral parameters alone. Each simulated spectrum is the spatial average of tissue-weighted spectral components over a user-selected volume of interest, where concentrations are sampled from literature-derived distributions and linewidths come from tissue-specific T2 relaxation times. Compared to 104 in-vivo spectra, 480 simulated spectra showed similar spectral shape and overlapping metabolite quantification, with comparable creatine SNR and linewidth; the simulations additionally covered parameter extremes such as high noise, lipid contamination, residual water, and shim broadening that the in-vivo set did not. The authors interpret this as evidence that the phantom is realistic enough for algorithm validation and flexible enough for augmentation.

Load-bearing premise

The evaluation assumes that tuning simulation parameters against the same in-vivo spectra used for comparison gives an unbiased measure of realism, rather than overfitting those specific spectra.

Editorial extensions

If this is right

Researchers can generate large, labeled MRS datasets with known ground-truth concentrations for training and validating quantification algorithms and machine-learning models.
The ability to tune SNR, linewidth, lipid contamination, residual water, and shim effects enables targeted robustness testing of processing pipelines.
Because simulations are saved in a standard open data format, they can be dropped into existing neuroimaging workflows without format conversion.
Simulated spectra include variability beyond the in-vivo sample, making the phantom a source of data augmentation for underrepresented spectral features.
Swapping in different anatomical templates, metabolite databases, or basis sets lets users extend the framework to new populations and acquisition protocols.

Reading between the lines

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

The realism claim would be stronger if the simulation parameters were optimized on one cohort and then confirmed on an independent cohort; the current evaluation uses the same in-vivo spectra both to set inputs and to judge agreement.
The t-SNE clusters of in-vivo spectra that the simulations miss are dominated by residual water features, suggesting the stochastic water model could be improved by fitting its parameter distributions to real data rather than uniform ranges.
Because the same quantification tool was used to derive simulation inputs and to score the outputs, part of the quantitative agreement may reflect a closed loop; having an independent quantifier assess the simulated spectra would be a sharper test.
If the phantom proves portable across sites and vendors, it could serve as a virtual trial platform for harmonizing MRS quantification before large multi-site studies are run.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the short version: the paper is a real contribution—an open, modular MRS simulation framework—but the abstract's realism claim is stronger than the evidence. The validation is partly circular: Section 2.4.2 takes Osprey-quantified concentrations from the 104 in-vivo spectra, uses them as simulation inputs, optimizes the remaining parameters against those same spectra, and then reports that the simulated spectra 'closely matched' the in-vivo data. That comparison can be tuned into agreement, and the quantification agreement is largely a round-trip test of the forward model plus Osprey. The paper doesn't report which parameters were optimized or their final values, so we can't judge whether the fitted settings are plausible or overfit.

What's genuinely new: no prior MRS digital phantom integrates anatomical templates, a literature-derived metabolite concentration database, and a modular signal model with GUI and NIfTI-MRS export. The individual components exist, but the integrated, extensible framework is new and fills a practical gap. The signal model is physics-based and clearly described—metabolite basis sets with T2 weighting, Voigt-profile macromolecules, five-exponential water, triglyceride lipid basis, shim maps, and noise. The code is on GitHub and the outputs are in a standard format. That's real reproducible work.

The soft spots beyond the circularity are minor. The MM model uses empirical scaling and doesn't fully match Osprey's five-component fit; the paper acknowledges this. Residual water and lipid variation are stochastic, which can miss structured real-world patterns, also acknowledged. Spatial metabolite heterogeneity is not modeled. These are future-work items, not fatal flaws. The paper's own Discussion is honest about remaining significant differences.

Bottom line: this is a paper for MRS method developers and machine-learning people who need ground-truth training data. It deserves a serious referee. Conditional acceptance is right, but the authors should (1) state which parameters were optimized and their final values, ideally with physiological plausibility checks, (2) add a held-out or independent validation set, or clearly label the current comparison as a calibration check rather than independent realism, and (3) soften the abstract's 'closely matched' claim. I'd cite it as a tool once those validation issues are addressed.

Referee Report

3 major / 6 minor

Summary. This paper presents a modular digital phantom framework for simulating single-voxel MR spectroscopy data. The framework combines an anatomical skeleton (MRiLab or BigBrain-MR), a literature-based metabolite concentration and T2 database, a basis-set-based metabolite signal model, and parametric models for macromolecules, water, lipids, shim broadening, and noise. Outputs are generated as component-wise spectra and exported in NIfTI-MRS format. The authors demonstrate parameter-dependent spectral variation and compare 480 simulated spectra against 104 in-vivo spectra from a published multisite study, quantifying similarity by visual comparison, t-SNE embedding, SNR/FWHM statistics, and Osprey-based metabolite quantification. The central claim is that the framework produces realistic MRS data suitable for algorithm validation and data augmentation.

Significance. The framework addresses a real need for flexible, reproducible simulation of MRS data. Its explicit separation of spectral components, configurable GUI, open-source Python implementation, and NIfTI-MRS export make it a practical contribution. If validated with an independent evaluation, it could support algorithm development, robustness testing, and machine learning research. The main weakness is that the realism validation is partly in-sample: in-vivo Osprey estimates are used as simulation inputs and the remaining parameters are optimized on the same in-vivo dataset, so the favorable agreement in the figures does not by itself establish independence or generalization. The spectral shape and variability comparisons are less directly forced and provide some genuine evidence of realism, but the quantitative claims need stronger support.

major comments (3)

[Section 2.4.2 and Section 3.3] The realism evaluation is in-sample. Section 2.4.2 states that metabolite concentrations from the in-vivo spectra were quantified with Osprey and used as simulation inputs, and that the 'remaining parameters are optimized to maximize similarity between the simulated and in-vivo spectra.' Because the same 104 spectra are then used to assess agreement in Figures 6-8, the reported matches in spectral shape, SNR (107±11 vs. 143±39), FWHM (6.32±0.63 vs. 6.34±1.21 Hz), and t-SNE overlap can reflect fitting to the evaluation data rather than independent realism. Please report the optimized parameter values, test generalization with a held-out set (e.g., split-half or leave-one-out), and clarify which comparisons are confirmatory versus exploratory.
[Abstract and Figure 8] The abstract's claim that simulated spectra 'closely matched' in-vivo data in 'metabolite quantification' is only partially supported by the results. Figure 8 shows statistically significant differences for multiple metabolites, macromolecules, and lipid components. Attributing these differences to the 'larger simulated sample size' (Section 4.2) is not a sufficient explanation: sample size affects test power, not the mean or median of the comparison. Please either use an equivalence-testing framework to support the 'closely matched' wording or temper the claim to distributional overlap.
[Section 2.4.2 and Section 4.2] The manuscript does not report which parameters were optimized, their final values, or whether these values are physiologically plausible. Without this information, the reader cannot distinguish between a model that captures the underlying biological variability and a model that merely overfits the noise of the 104 acquisitions. This is especially relevant for noise_level, mm_level, lipid_amp_factor, water_amp_factor, and shim parameters, which are all free in Table 1. Providing the final optimized settings and a sensitivity analysis would strengthen the claim that the phantom is useful for augmenting new acquisitions.

minor comments (6)

[Section 2.3.1] The sentence 'The values for c_kℓ are obtained by random sampling from a Gaussian distribution defined by the mean and standard deviation provided in the metabolite dataframe. for' contains an orphan 'for' after the sentence period; please remove or complete it.
[Section 3.2] The text says the example basis set is for sLASER at TE 30 ms, while Section 2.3.1 states that automated basis-set generation currently supports PRESS. Clarify whether this example used a pre-generated basis set, since this is potentially confusing for users.
[Section 2.3.4, Eq. (10)] The notation in the lipid signal model contains garbled sub/superscripts (FℱG 1H<A FID1(t)). Please typeset the equation cleanly.
[Reference 36] The URL for van der Maaten's t-SNE paper ends with '?fbcl', which appears to be a corrupt tracking parameter; please replace with the canonical URL.
[Section 2.3.5] The factor 10^3 in Eq. (11) is described as 'ensuring realistic noise amplitudes,' but its origin is not explained. A sentence clarifying how this factor was determined would improve reproducibility.
[Section 2.2.1] The description of filtering for the T2 database says 'The T2 database is filtered on studies that use 3T scanners,' but the relation between the T2 filter and the T1 placeholder in Table S1 is not stated. If T1 values are placeholders, note this explicitly in the main text.

Circularity Check

2 steps flagged · score 6.0 of 10

In-vivo realism evaluation is in-sample: Osprey quantification outputs are fed into the simulator and later compared with the same in-vivo spectra, while remaining simulation parameters are optimized against the same evaluation set.

fitted input called prediction [Section 2.4.2, Comparison with In-Vivo Data]
"Metabolite concentrations from these in-vivo spectra are quantified using the Osprey toolbox, and the resulting estimates serve as input for the phantom to generate simulated spectra that closely mimic the corresponding in-vivo acquisitions."

The simulation's tissue-specific metabolite concentrations are taken from Osprey quantification of the very in-vivo spectra used as the comparison target. Later, Section 3.3 reports that simulated spectra are quantified with Osprey and the results are compared with the in-vivo values (Figure 8). Re-quantifying simulated spectra built from those same Osprey estimates primarily tests round-trip consistency of the forward model and basis set, not biological realism. The 'closely matched' metabolite quantification claim is therefore forced by construction, since the input concentrations are the evaluation target.
fitted input called prediction [Section 2.4.2, Comparison with In-Vivo Data]
"The simulations replicate the acquisition and sequence parameters of the in-vivo dataset, and the remaining parameters are optimized to maximize similarity between the simulated and in-vivo spectra."

The paper states that the free simulation parameters (noise level, MM scaling, lipid and water amplitudes, shim settings, etc.) are optimized against the same 104 in-vivo spectra later used for the realism comparison. Consequently, the reported spectral-shape overlap, t-SNE closeness, SNR (107 +/- 11 versus 143 +/- 39), and FWHM (6.32 +/- 0.63 versus 6.34 +/- 1.21 Hz) are in-sample fit diagnostics rather than independent out-of-sample evidence. The optimized parameter values are not reported and no held-out acquisition is tested, so the match does not demonstrate that a fixed parameter set generalizes to new data.

full rationale

The construction of the phantom itself is largely independent: the skeleton is anatomical, the metabolite and T2 values come from a literature meta-analysis, the basis sets are generated externally, and the signal model is physics-based. These parts are not circular. However, the central realism evaluation is partially self-fulfilling. The quantitative validation is circular because Osprey estimates of the in-vivo spectra are used as simulation inputs and then re-quantified as the comparison target. The spectral-shape and variability validation is weakened because the remaining simulation parameters were explicitly optimized to maximize similarity to the same in-vivo dataset, and no held-out data are used. The self-citations to the authors' meta-analysis and to MRSCloud are not load-bearing circularity because they supply external literature data and tools rather than the conclusion of realism. Overall, the framework has independent content, but the headline claim of 'closely matched' in-vivo realism is supported by an in-sample, partly by-construction evaluation, so a score of 6 is appropriate.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The framework's realism depends on a stack of literature-derived models: meta-analysis concentrations, MRSCloud basis sets, Wright MM model, Lin water model, and SimnTG lipid spin systems. These are domain assumptions with internal checks but no independent validation in this paper. The only fitted parameters explicitly mentioned are the 'optimized' tuning parameters in the in-vivo comparison and the empirically tuned MM amplitude.

free parameters (2)

optimized simulation parameters = unspecified
Section 2.4.2: 'remaining parameters are optimized to maximize similarity between the simulated and in-vivo spectra.' These include noise level, MM scaling, and possibly water/lipid parameters; since they are tuned on the same in-vivo dataset used for evaluation, the reported similarity is inflated.
mm_level (global macromolecule scaling) = 25.0 in example config
Section 3.3 / Figure S2: MM amplitude empirically tuned; Discussion says 'empirical tuning of MM amplitudes in the simulation' contributes to quantification differences.

assumptions (6)

domain assumption Metabolite concentrations and T2 relaxation times from the Gudmundson meta-analysis (ref 19) are representative of healthy adult brain at 3T.
Section 2.2.1: The entire phantom's metabolite content is built from this database after filtering; if the meta-analysis is biased, all simulated spectra inherit the bias.
domain assumption Basis sets generated by MRSCloud with universal pulses accurately represent true metabolite signals for the chosen localization and TE.
Section 2.3.1: The framework accepts or auto-generates basis sets; the realism of metabolite simulation depends on basis set accuracy.
domain assumption The macromolecule model of Wright et al. (ref 25) with 3T parameters from Landheer, Hoefemann, and Hupfeld produces a realistic MM background.
Section 2.3.2: MM component linewidths, amplitudes, and relaxation times are taken from published experimental studies; the authors note tissue-specific MM differences are not yet available at 3T.
domain assumption The residual water model based on Lin et al. (ref 30) with five damped exponentials represents in-vivo water signals.
Section 2.3.3: Water signal composition is adopted from prior work without re-validation in this paper.
domain assumption The assumption that human white adipose tissue >90% consists of oleic, palmitic, linoleic, and palmitoleic acids (Hodson et al.) is sufficient for lipid contamination.
Section 2.3.4: Triglyceride spin systems built from these four fatty acids; if other fatty acids contribute significantly, lipid contamination shape could differ.
standard math The Voigt linewidth approximation of Olivero and Longbothum (Eq. 5) is valid for the MM components.
Section 2.3.2: Used to derive Gaussian linewidths from measured FWHM and T2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Digital Phantom for MR Spectroscopy Data Simulation." pith.science (2026). https://pith.science/paper/HFCC5BWJ

@misc{pith2026241215869,
  author       = {Pith},
  title        = {Pith review of: A Digital Phantom for MR Spectroscopy Data Simulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HFCC5BWJ}},
  note         = {Machine review of arXiv:2412.15869}
}

read the original abstract

Simulated data is increasingly valued by researchers for validating MRS processing and analysis algorithms. However, there is no consensus on the optimal approaches for simulation models and parameters. This study introduces a novel MRS digital brain phantom framework, providing a comprehensive and modular foundation for MRS data simulation. The framework generates a digital brain phantom by combining anatomical and tissue label information with metabolite data from the literature. This phantom contains all necessary information for simulating spectral data. The MRS phantom is combined with a signal-based model to demonstrate its functionality and usability in generating various spectral datasets. Outputs can be saved in the NIfTI-MRS format, enabling their use in downstream applications. To evaluate the realism of the simulated spectra, a comparison was performed against in-vivo MRS data acquired under similar conditions. The phantom was implemented using two anatomical templates at different resolutions and tested across a range of user-defined simulation parameters. Simulated spectra exhibited realistic signal characteristics and structural variability. When compared to in-vivo data, the simulated spectra closely matched in terms of spectral shape, signal-to-noise ratio, and metabolite quantification. The simulations also captured key variability features and provided additional diversity not present in the in-vivo dataset, supporting use in robustness testing and data augmentation. This novel digital phantom provides a flexible and extensible platform for MRS data simulation. Its modular architecture, user-friendly GUI, and open-source implementation support reproducible research, algorithm development, and validation in the MRS community.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[13]

Fast Realistic MRI Simulations Based on Generalized Multi-Pool Exchange Tissue Model

Liu F , Velikina JV , Block WF , Kijowski R, Samsonov AA. Fast Realistic MRI Simulations Based on Generalized Multi-Pool Exchange Tissue Model. IEEE Trans Med Imaging. 2017;36(2):527-537. doi:10.1109/TMI.2016.2620961 14. C T, C L, Lr S, Fg Z. VirtMRI: A Tool for Teaching MRI. J Med Syst. 2023;47(1). doi:10.1007/s10916-023-02004-4 15. Stöcker T, Vahedipour...

arXiv 2017
[25]

Relaxation-corrected macromolecular model enables determination of 1H longitudinal T1-relaxation times and concentrations of human brain metabolites at 9.4T

Wright AM, Murali-Manohar S, Borbath T, Avdievich NI, Henning A. Relaxation-corrected macromolecular model enables determination of 1H longitudinal T1-relaxation times and concentrations of human brain metabolites at 9.4T. Magn Reson Med. 2022;87(1):33-49. doi:10.1002/mrm.28958 26. Landheer K, Gajdosik M, Treacy M, Juchem C. Concentration and eDective T-2...

work page doi:10.1002/mrm.28958 2022
[37]

skeleton

Degaonkar MN, Pomper MG, Barker PB. Quantitative proton magnetic resonance spectroscopic imaging: Regional variations in the corpus callosum and cortical gray matter. J Magn Reson Imaging. 2005;22(2):175-179. doi:10.1002/jmri.20353 38. Emir UE, Auerbach EJ, Moortele PFVD, et al. Regional neurochemical proﬁles in the human brain measured by 1H MRS at 7 T u...

work page doi:10.1002/jmri.20353 2005

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [13]

Fast Realistic MRI Simulations Based on Generalized Multi-Pool Exchange Tissue Model

Liu F , Velikina JV , Block WF , Kijowski R, Samsonov AA. Fast Realistic MRI Simulations Based on Generalized Multi-Pool Exchange Tissue Model. IEEE Trans Med Imaging. 2017;36(2):527-537. doi:10.1109/TMI.2016.2620961 14. C T, C L, Lr S, Fg Z. VirtMRI: A Tool for Teaching MRI. J Med Syst. 2023;47(1). doi:10.1007/s10916-023-02004-4 15. Stöcker T, Vahedipour...

arXiv 2017

[2] [25]

Relaxation-corrected macromolecular model enables determination of 1H longitudinal T1-relaxation times and concentrations of human brain metabolites at 9.4T

Wright AM, Murali-Manohar S, Borbath T, Avdievich NI, Henning A. Relaxation-corrected macromolecular model enables determination of 1H longitudinal T1-relaxation times and concentrations of human brain metabolites at 9.4T. Magn Reson Med. 2022;87(1):33-49. doi:10.1002/mrm.28958 26. Landheer K, Gajdosik M, Treacy M, Juchem C. Concentration and eDective T-2...

work page doi:10.1002/mrm.28958 2022

[3] [37]

skeleton

Degaonkar MN, Pomper MG, Barker PB. Quantitative proton magnetic resonance spectroscopic imaging: Regional variations in the corpus callosum and cortical gray matter. J Magn Reson Imaging. 2005;22(2):175-179. doi:10.1002/jmri.20353 38. Emir UE, Auerbach EJ, Moortele PFVD, et al. Regional neurochemical proﬁles in the human brain measured by 1H MRS at 7 T u...

work page doi:10.1002/jmri.20353 2005