REVIEW 3 major objections 2 minor 3 references
MIMOSA: Multi-parametric Imaging using Multiple-echoes with Optimized Simultaneous Acquisition for highly-efficient quantitative MRI
T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read MIMOSA claims that a single 3-minute, 1-mm whole-brain scan can produce accurate T1, T2, T2*, proton density, and source-separated susceptibility maps, with 11.8-fold acceleration and high scan-rescan repeatability.
desk verdict Promising abstract, but the attached full text is a different paper, so there is no method to check. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the combined acquisition/reconstruction chain: a single interleaved module that appends a multi-echo gradient-echo block to a 3D Look-Locker FLASH readout with T2 preparation, sampled with a spiral-like Cartesian trajectory, and reconstructed by a multi-contrast zero-shot self-supervised network. The sequence design determines how T1, T2, and T2*-weighted contrasts are co-encoded in one k-space data set; the reconstruction network, trained without external ground truth on the same undersampled data, separates those encodings into the five parameter maps. Simulation-based sequence optimization and phantom/reference validation are what convert the raw contrast sep
What would settle it
Run MIMOSA on an ISMRM/NIST phantom at accelerations R = 3.3, 6.5, and 11.8 and compare each estimated parameter with the known reference values; if error grows sharply with R, or if volunteer parameter values shift systematically between acceleration factors, the accuracy claims fail.
Extended reading notes
Core claim
The central discovery is a pulse sequence, MIMOSA, that encodes five tissue parameters in one interleaved acquisition. It extends 3D-QALAS by appending a multi-echo gradient-echo module to the T2-prepared Look-Locker 3D turbo FLASH readout and sampling k-space in a spiral-like Cartesian order, then reconstructs all contrasts with a multi-contrast zero-shot self-supervised network. In simulations the design improves parameter estimation accuracy over 3D-QALAS; in ISMRM/NIST phantom and in-vivo experiments it matches reference methods, and scan-rescan reproducibility is high (ICC up to 0.998). The practical claim is that whole-brain 1-mm isotropic T1, T2, T2*, PD, and source-separated QSM can
Load-bearing premise
The single joint signal model connecting the interleaved acquisitions to T1, T2, T2*, PD, and susceptibility must be correct and unbiased at every acceleration factor, including the highest tested value of R = 11.8.
Editorial extensions
If this is right
- Whole-brain protocols could collapse five separate quantitative scans (T1, T2, T2*, PD, QSM) into one 3-minute acquisition, reducing patient time and motion sensitivity.
- Acceleration beyond R = 11.8 may become practical if the zero-shot reconstruction remains stable, making even faster or higher-resolution protocols possible.
- Source-separated paramagnetic and diamagnetic susceptibility maps from a routine scan would let clinicians and researchers dissect iron, myelin, and calcification contrasts without extra scan time.
- The 7T 750-micron acquisition opens mesoscale quantitative mapping of cortical layers and subcortical structures in 13 minutes.
- The high scan-rescan ICCs support longitudinal monitoring, where stable repeatability matters more than cross-sectional accuracy.
Reading between the lines
- A direct test beyond the reported experiments would be to measure the same phantom at R = 11.8 and compare per-parameter errors with R = 3.3, since phantom validation may be reported at lower accelerations.
- If the underlying joint signal model is unbiased, the same acquisition skeleton could be extended to other contrasts (e.g., myelin water fraction or CEST) by swapping the encoding module and re-running the simulation-based optimization.
- The 7T mesoscale maps should be validated against histology or post-mortem iron measurements, because scan-rescan repeatability alone does not establish biological accuracy at 750 microns.
- Portability of the method across field strengths and B1+ inhomogeneities would require explicit B1+ mapping or calibration in the reconstruction; the abstract does not state that such corrections were applied.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The submission (arXiv:2508.10184) is presented as a physics/medicine manuscript, with an abstract describing MIMOSA, a new MRI sequence for simultaneous T1, T2, T2*, PD, and source-separation QSM mapping. The abstract reports simulation-based optimization, phantom and in-vivo validation against 3D-QALAS and reference techniques, acceleration factors up to R = 11.8, scan-rescan ICCs between 0.947 and 0.998, and scan times of 3 min at 3T and 13 min at 7T. However, the supplied full text is arXiv:2508.10180v3, an unrelated cs.CL paper titled 'For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs.' No MIMOSA sequence description, signal model, reconstruction details, simulation setup, phantom results, in-vivo statistics, or analysis code is present. Consequently, the abstract's quantitative claims cannot be inspected or verified from the submitted manuscript.
Significance. If substantiated, the MIMOSA contribution would be significant: simultaneous high-resolution multi-parametric mapping with acceleration up to R = 11.8 and good repeatability would be of strong interest to the quantitative MRI community. The abstract's design—comparison against external reference techniques and ISMRM/NIST phantom data—is a non-circular evaluation strategy, and the use of scan-rescan ICCs is appropriate for repeatability assessment. However, significance cannot be assessed from the submitted text because none of the supporting methods or data for MIMOSA are present. The supplied full text provides no evidence for the central claims and, in fact, describes an entirely different research area.
major comments (3)
- [Full text (all sections)] The submitted full text is not the MIMOSA manuscript. It is arXiv:2508.10180v3, 'For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs,' a cs.CL paper whose abstract, introduction, method, experiments, limitations, and appendix all concern data valuation for large language models and vision-language models. None of the MIMOSA claims can be checked: there is no pulse sequence diagram, no signal equation, no flip-angle or echo-time schedule, no B1+/off-resonance correction description, no zero-shot self-supervised reconstruction formulation, no phantom results table, and no in-vivo reconstruction details. This is a load-bearing absence because the abstract's accuracy, ICC, and acceleration claims cannot be validated from any text in the submission.
- [Abstract, Results] The abstract reports scan-rescan ICCs of 0.998 (T1), 0.973 (T2), 0.947 (T2*), 0.992 (QSM), 0.987 (paramagnetic), and 0.977 (diamagnetic), and states that MIMOSA showed 'better agreement with reference techniques than 3D-QALAS' in phantom experiments. No sample sizes, confidence intervals, p-values, or effect sizes are provided. While abstracts often omit such details, the complete absence of the corresponding methods and results sections in the submission makes these numbers uninterpretable and unverifiable.
- [Abstract, Methods] The central claim that a single MIMOSA acquisition yields accurate T1, T2, T2*, PD, and QSM maps depends on the correctness and unbiasedness of the joint signal model and the reconstruction. The abstract states only that simulations were performed to optimize the sequence and that a zero-shot self-supervised learning algorithm was used. No model equation is given, and no analysis of acceleration-dependent bias is provided. In particular, the transfer of accuracy from lower acceleration factors (R = 3.3, 6.5) to R = 11.8, and from 3T to 7T at 750 um, is asserted without any inspectable evidence. This is a load-bearing omission, not a stylistic one.
minor comments (2)
- [Abstract] The abstract uses 'source separation QSM,' 'mesoscale quantitative mapping,' and 'spiral-like Cartesian trajectory' without definitions; a methods section would be needed to clarify these terms. Also, the acronym MIMOSA is expanded only in the title, and the abstract would benefit from stating the acquisition time separately from the reconstruction time.
- [General] The submission's own limitation section (Section 7 of the supplied full text) discusses limitations of forward-only data valuation for LLMs and VLMs. That limitation statement explicitly refers to the wrong manuscript and cannot serve as a limitation disclosure for MIMOSA. The authors should ensure that the correct manuscript, including its limitations, is submitted.
Circularity Check
No circular step is demonstrable; the attached body text is an unrelated cs.CL paper, so the MIMOSA claims are unverifiable but not shown to be circular.
full rationale
The abstract validates MIMOSA against external benchmarks—ISMRM/NIST phantom, reference techniques, and scan-rescan ICCs—which is the structurally non-circular form of evaluation. No fitted parameter is relabeled as a prediction: the R = 3.3/6.5/11.8 acceleration conditions are tested against measured maps, and the reported ICCs are independent repeatability statistics. However, the supplied full text is not the MIMOSA manuscript; it is arXiv:2508.10180 (For-Value, a cs.CL data-valuation paper), so the actual signal model, simulation cost function, and reconstruction loss are not available for inspection. Without equations or an explicit model, no identity of the form 'Eq. X = Eq. Y by construction' or 'fitted input renamed as prediction' can be exhibited. Per the hard rules, circularity cannot be inferred from missing evidence alone; the mismatch is a manuscript-integrity and verifiability concern, not a demonstrated circular step. Therefore the appropriate circularity score is 0, with no circular steps identified.
Assumptions & free parameters
free parameters (2)
- MIMOSA sequence timing parameters (T2 prep durations, flip angle schedule, echo times, number of echoes) =
not stated in abstract
- Zero-shot self-supervised reconstruction hyperparameters and learned network weights =
not stated in abstract
assumptions (3)
- domain assumption The Bloch equation signal model of interleaved Look-Locker recovery with T2 preparation and multi-echo FLASH readout predicts the measured signal for each tissue
- domain assumption Mono-exponential relaxation and standard QSM dipole inversion and source separation models hold for brain tissue
- domain assumption The zero-shot self-supervised reconstruction yields unbiased parameter estimates at high acceleration
Cite this review
Pith. "Pith review of MIMOSA: Multi-parametric Imaging using Multiple-echoes with Optimized Simultaneous Acquisition for highly-efficient quantitative MRI." pith.science (2026). https://pith.science/paper/UYMB2P5K
@misc{pith2026250810184,
author = {Pith},
title = {Pith review of: MIMOSA: Multi-parametric Imaging using Multiple-echoes with Optimized Simultaneous Acquisition for highly-efficient quantitative MRI},
year = {2026},
howpublished = {\url{https://pith.science/paper/UYMB2P5K}},
note = {Machine review of arXiv:2508.10184}
}
read the original abstract
Purpose: To develop a new sequence, MIMOSA, for highly-efficient T1, T2, T2*, proton density (PD), and source separation quantitative susceptibility mapping (QSM). Methods: MIMOSA was developed based on 3D-quantification using an interleaved Look-Locker acquisition sequence with T2 preparation pulse (3D-QALAS) by combining 3D turbo Fast Low Angle Shot (FLASH) and multi-echo gradient echo acquisition modules with a spiral-like Cartesian trajectory to facilitate highly-efficient acquisition. Simulations were performed to optimize the sequence. Multi-contrast/-slice zero-shot self-supervised learning algorithm was employed for reconstruction. The accuracy of quantitative mapping was assessed by comparing MIMOSA with 3D-QALAS and reference techniques in both ISMRM/NIST phantom and in-vivo experiments. MIMOSA's acceleration capability was assessed at R = 3.3, 6.5, and 11.8 in in-vivo experiments, with repeatability assessed through scan-rescan studies. Beyond the 3T experiments, mesoscale quantitative mapping was performed at 750 um isotropic resolution at 7T. Results: Simulations demonstrated that MIMOSA achieved improved parameter estimation accuracy compared to 3D-QALAS. Phantom experiments indicated that MIMOSA exhibited better agreement with the reference techniques than 3D-QALAS. In-vivo experiments demonstrated that an acceleration factor of up to R = 11.8-fold can be achieved while preserving parameter estimation accuracy, with intra-class correlation coefficients of 0.998 (T1), 0.973 (T2), 0.947 (T2*), 0.992 (QSM), 0.987 (paramagnetic susceptibility), and 0.977 (diamagnetic susceptibility) in scan-rescan studies. Whole-brain T1, T2, T2*, PD, source separation QSM were obtained with 1 mm isotropic resolution in 3 min at 3T and 750 um isotropic resolution in 13 min at 7T. Conclusion: MIMOSA demonstrated potential for highly-efficient multi-parametric mapping.
Reference graph
Works this paper leans on
-
[1]
For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs Wenlong Deng1⋆, Qi Zeng3, Jiaming Zhang1, Minghui Chen1, Zixin Ding3, Christos Thrampoulidis1,Boying Gong 3†,Xiaoxiao Li 1,2† 1University of British Columbia, 2Vector Institute, 3Meta ⋆Work done at Meta, †Corresponding author Abstract Data valuation is essential for enhancing th...
work page 2023
-
[7]
A.6.1 Baseline Checkpoints Selection For baseline methods, we select the model check- point with the highest test AUC, as influence function-based methods exhibit significant perfor- mance variability across training checkpoints. No- tably, this variability does not correlate with vali- dation loss, posing challenges for practical deploy- ment. We compare...
work page 2022
-
[2024]
For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs
and Shapley value-based methods (Ghorbani and Zou, 2019), which provide frameworks for estimating how individual data points affect model predic- tions (Kwon et al., 2024; Zhou et al., 2024). These methods have proven effective in downstream ap- plications such as detecting mislabeled data (Koh and Liang, 2017; Kwon et al., 2024), identifying influential ...
work page Pith review arXiv 2019
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.