List of Papers Browse by Subject Areas Author List
Abstract
Digital phantoms play a key role in MRI for pulse sequence optimization, artifact correction, and clinical trials, as well as for generating physically grounded synthetic data via augmentation. A digital phantom is defined by quantitative parameter maps, including longitudinal relaxation time (T1), transverse relaxation time (T2), and proton density (PD). We propose a self-supervised, flow-based deep learning framework for estimating quantitative T1, T2, and PD maps from conventional weighted brain MR images. The method explicitly incorporates MRI pulse sequence parameters—repetition time (TR), echo time (TE), and echo spac-ing (ES)—allowing the model to better capture the underlying MR signal formation process while preserving high anatomical fidelity. The model takes as input T1-, T2-, and PD-weighted images together with their corresponding sequence parameters and outputs quantitative maps, which are subsequently used by an MR signal model to synthesize new weighted images. The framework is trained on a combination of synthetic and real Turbo Spin Echo (TSE) data with varying TR, TE, and ES values. We evaluated different strategies for integrating MRI pulse sequence parameters into the flow-based model. Quantitative evaluation demonstrates high similarity between original and synthesized images, achieving MS-SSIM scores of 0.9915, 0.9464, and 0.9771 for T1-, T2-, and PD-weighted images, respectively. Source code is available on GitHub https://github.com/KseniyaMD/flow_based_phantoms
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2599_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/KseniyaMD/flow_based_phantoms
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BelKse_SequenceConditioned_MICCAI2026,
author = { Belousova, Kseniya AND Badrieva, Zilya AND Brui, Ekaterina AND Al-Haidri, Walid},
title = { { Sequence-Conditioned Flow-Based Models for Digital Phantom Generation in MRI } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This work presents a MRI scan parameter-conditioned, flow-based deep learning model for joint mapping of T1, T2, and PD parameters. A hierarchy flow model is trained on T1, T2, and PD weighted images, optionally with corresponding scan parameters (TR, TE, and ESP) of TSE imaging. Three different approaches to integrating these MRI sequence parameters into the baseline model were compared on both synthetic images and real-world datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
No need for ground-truth parametric maps. Only one set of synthetic quantitative maps is needed.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Novelty of the model architecture is moderate-to-weak. The current network is built on existing Hierarch Flow with marginal modifications.
- The physical model (Eq. 1) does not take a stimulated echo (caused by an imperfect 180 degree refocusing RF) into account, nor the T2 weighting along the echo-train.
- Limited imaging scenarios tested. Only TSE imaging with tight ranges of the scan parameters. A simpler SE imaging could have been tested to better describe Eq. 1.
- Limited evaluation. Only normal tissues were considered. Only SSIM and PSNR were reported. Quantification errors need to be reported, at least for the synthetic datasets. No comparison against other SOTA methods
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The method’s novelty is rather moderate, and there are several limitations as noted above. But this paper may pull attention from the quantitative MRI folks.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors’ feedback is mostly acceptable. This paper will pull attention from quantitative MRI folks, and thus deserves a room for discussion.
Review #2
- Please describe the contribution of the paper
A hierarchical flow-based self-supervised learning deep learning framework to synthesize quantitative maps (T1, T2, and PD) from T1-, T2-, and PD-weighted conventional brain MRI images, which are subsequently leveraged to generate new synthetic brain images.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The content is well organized without grammatical errors or typos.
- The content clearly communicates the idea
- The synthetic dataset generation method is sensible
- Figures are well-stylized and legible
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- Scaling intensities to a maximum of 4096 prior to loss calculation is not justified.
- Parameter counts comparisons of the three integration strategies are not presented. For example, what are the training, memory, and inference overheads of leveraging the Sequence-Aware Affine Network approach?
- The authors are highly encouraged to evaluate the quality of the synthesized magnitude images on a downstream task. For example, compare performances on brain segmentation between a network trained only on real data vs the same network trained on real + synthesized data. Does the second network achieve higher scores? Is it more robust? Or can it achieve the same performance as the first network but with limited real data (because the synthetic data makes up for the scarcity of real data)?
- Training dynamics are not justified. How was the LR chosen? Why 17 + 50 epochs? Why not early stopping? What does “sufficient to observe full convergence” mean, i.e. what MSE was considered full convergence?
- What was the wall clock duration for training?
- In Table 1, why are all the values in the EF and SA columns highlighted in bold since they are not the same values in all instances? Also, why are the values reported only up to a precision of 2, this is insufficient.
- No comparisons with baselines are provided (even if comparisons are restricted to methods that can only generate single maps)
- The authors are also recommended to comment about how this work can be adapted to pulse sequences beyond TSE.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work is valuable, however the authors have not thoroughly explored and evaluated their contribution. Further, the submission lacks several details that affect reproducibility in the event that the source code is not published.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors are commended for their diligent responses to my feedback. They have sufficiently addressed my queries and will hopefully include the same information in the final edit.
Review #3
- Please describe the contribution of the paper
The paper describes a method for the synthesis of phantom MRI images by firstly deriving the quantitative maps of some other initial weighted MRI scans. The images are first transformed to a latent space where features can then be used to inform a style network that is then in charge of transforming these features into the target domain, in this case, the quantitative maps. The network is informed by the acquisition settings as well as features from the quantitative maps, allowing for the network to transform the latent representation of the source images, the weighted MRIs, into the target domain. The output quantitative maps are then transformed back to the original weighted images, allowing for training via error between the input and its synthesised version.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper offers a novel approach for the generation of synthetic weighted MRI scans, allowing for the specification of certain acquisition settings. The injection of the style features extracted from the quantitative maps sounds interesting as it does not require the pairing of weigthed and quantitative MRI images.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper provides a thorough evaluation of their approach under different settings, however, it lacks of comparisons to other methods, such as MRI2Qmap (Golbabaee, et. al., 2024, with an in-depth arxiv version already out), while the paper itself might differ, it could be adapted and serve as a point of comparison.
The metrics are reported on the original vs synthesized MRI images, however, as this paper also concerns with the generation of new data, reporting other metrics such as the Fréchet Inception Distance (FID) could show this aspect of the proposed work. I would also be interested in seeing metrics regarding the generation of the quantitative maps and their closeness to the GT maps used to extract the features. For instance, generate weighted images from the quantitative maps (not used during training) and use them as input to the network and check the similarity of the quantitative maps generated by them against the original quantitative scans.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Despite the weaknesses identified, the contribution of this paper is well founded and a valuable piece of work that could help not only in the generation of MRI datasets, but it could also address the task of the reconstruction/estimation of quantitative MRI maps, which itself broadens the applications and usefulness of the paper.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I do believe the paper offers a novel approach, however, evaluation of the method against other works found in the literature could have further elevated the paper. While the mertis are still present, lacking of metrics to fully assess the results as well as other works can affect the perception of the advancement in the field.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The presented paper provides a valuable approach that addresses both qMRI and synthetic data generation that is worth disseminating. Authors clarify this is a proof of concept that can be further explored given the promising results presented here, and are aware of the limitations of the work – this should be made clear in the final version of the manuscript.
Author Feedback
R1: Novelty of the model architecture Our contribution is methodological rather than purely architectural. Although Hierarchy Flow serves as the backbone, the proposed Sequence-Aware Affine Network (SA) introduces a physics-driven multi-channel conditioning mechanism. SA processes each contrast channel (T1w, T2w, PDw) through dedicated sub-networks that learn non-linear relationships between pulse-sequence parameters (TR/TE/ES) and signal intensity. We will clarify this distinction in the Introduction. R1: The physical model We agree that the current MR signal model does not account for RF pulse profiles, stimulated echoes, or mixed contrast effects caused by turbo-spin echo acceleration, T2 decay, and phase-encoding order. However, this study focuses on feasibility evaluation for self-supervised phantom generation. The analytical approximation remains suitable for TSE sequences because image contrast is mainly determined by the effective echo filling the central k-space. We agree that more advanced modeling is required for greater robustness, and future work will incorporate the MRZero simulator. R2: Adaptation to other pulse sequences The framework is generalizable to pulse sequences with known analytical signal equations, provided the inputs contain distinguishable T1, T2, and PD contrast information. R2: Scaling, parameter counts, and training dynamics Scaling to 4096 corresponds to the 12-bit dynamic range of DICOM MRI. Compared with normalization to [0,1], this scaling avoids vanishingly small losses and improves numerical stability in flow-based optimization. Model complexity: Baseline: 1.02M params / 64 ms inference / 11.96 MB memory; EF: 1.04M / 65 ms / 12.18 MB; FF: 1.08M / 74 ms / 12.71 MB; SA: 1.06M / 215 ms / 12.51 MB. Training on an NVIDIA RTX 5070 (12 GB VRAM) required ~80 minutes for all models. The learning rate and “17 + 50 epochs” schedule were selected empirically from pilot experiments and loss curves. Convergence was defined as a loss change <0.0001 over five consecutive epochs. R2: Downstream task evaluation Evaluation on downstream tasks (e.g., segmentation trained on synthetic vs. real data) is an important next step but falls outside the scope of the current submission. Conference guidelines also prevent adding new experiments. R2: Table 1 formatting and precision Highlighting both EF and SA was a formatting mistake. With 4-decimal precision, MS-SSIM on the real test dataset is: Baseline 0.9825/0.9386/0.9640, FF 0.9844/0.9436/0.9682, EF 0.9897/0.9360/0.9812, SA 0.9915/0.9464/0.9771 for T1/T2/PD, respectively. R1, R3: Limited evaluation and imaging scenarios Healthy volunteer data was used to validate technical feasibility; evaluation on pathological open datasets is planned as immediate future work. We report MS-SSIM and PSNR on synthesized weighted images because the primary objective is anatomically faithful MRI augmentation rather than direct quantitative-map estimation. Quantitative maps were verified against ground truth on synthetic data during development, but submission rules prohibit adding these results. Since this work is a feasibility study, we selected representative pulse sequences. While SE sequences are easier to approximate analytically, they are less common clinically. R1, R2, R3: Comparisons with other methods Direct comparison with prior approaches is difficult because existing methods address different constraints. Our framework simultaneously generates three quantitative maps from conventional weighted MRI in a self-supervised manner without ground-truth data, while explicitly incorporating pulse-sequence parameters and maintaining efficient training and inference. To our knowledge, none of the reviewed methods satisfy all these conditions simultaneously. Regarding MRI2QMap, the method was designed for MR fingerprinting and highly undersampled 3D MRI acquired with a TGAS–SPI–MRF sequence. In contrast, our approach targets standard pulse sequences available on clinical scanners.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper proposes a sequence-conditioned hierarchical flow-based framework for digital phantom generation in MRI. The model jointly infers T1, T2, and PD quantitative maps from weighted MRI (T1w, T2w, PDw) without requiring paired ground-truth qMaps. The reviewers agree that the pairing-free formulation and sequence-conditioned design are practically valuable, that the writing and figures are clear, and that the problem has broad potential impact for qMRI reconstruction and synthetic MRI dataset generation.
Scores split 3/4/4.The AC finds that R1 provides the deepest concerns. Eq. 1 omits stimulated echoes from imperfect 180-degree refocusing and T2 weighting along the TSE echo train, which is a non-trivial physics omission; evaluation is narrow (only normal tissue, only SSIM/PSNR, no quantification-error reporting, no SOTA comparisons); and testing is confined to TSE with tight parameter ranges. R2 adds concrete reproducibility concerns — unjustified LR and epoch counts, undefined convergence criterion, no wall-clock time, unjustified intensity scaling to 4096, missing parameter/memory comparisons across the three integration strategies, Table 1 formatting issues. R3 asks for baseline positioning against MRI2Qmap (Golbabaee et al. 2024), FID-style generation metrics, and a round-trip qMap evaluation. On the weaknesses, the AC also finds R1 and R3 disagree on novelty framing. R1 views the architecture as Hierarchical Flow with marginal changes; R3 treats the sequence-conditioned phantom-generation formulation as novel.
Some of the concerns are clarification-tractable (reproducibility, parameter counts, intensity scaling, physical-model assumptions), while the deepest ones (correcting the physical model, adding SOTA comparisons, downstream-task evaluation, and FID / quantification-error reporting) would require new experiments. R1 explicitly notes that some concerns are not rebuttal-tractable. Given the slightly positive score and the mix of tractable and non-tractable concerns, I recommend inviting the authors to respond in rebuttal before a final decision, with attention to reproducibility, the physical-model assumptions, and baseline positioning.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The AC thanks the authors and all three reivewers for the active engagement in the rebuttal phase. The original review raised concerns regarding methodological novelty, underlying running hypotheses, and baseline positioning. The authors have effectively addressed all these concerns and acknowledged potential limitations that are infeasible to be addressed in a clarification-focused rebuttal, and clarified that the paper should be viewed as a proof of novel concept . All three reviewers agreed that the paper merits acceptance after rebuttal. Considering the reviewers’ positive score and the merits of the work, I recommend acceptance.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers unanimously recommend this paper for acceptance based on the rebuttal and note no specific unresolved concerns. The authors should be careful to include known limitations in the final manuscript. Based on this, I have no problem recommending this paper for acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers now recommend clear acceptance, highlighting the substantial interest in quantitative MRI and methodological novelty through the injection of new features not requiring paired data.
