Abstract

Segmenting stroke lesions in MRI is challenging due to diverse acquisition protocols that limit model generalisability. In this work, we introduce two physics-constrained approaches to generate synthetic quantitative MRI (qMRI) images that improve segmentation robustness across heterogeneous domains. Our first method, qATLAS, trains a neural network to estimate qMRI maps from standard MPRAGE images, enabling the simulation of varied MRI sequences with realistic tissue contrasts. The second method, qSynth, synthesises qMRI maps directly from tissue labels using label-conditioned Gaussian mixture models, ensuring physical plausibility. Extensive experiments on multiple out-of-domain datasets show that both methods outperform a baseline UNet, with qSynth notably surpassing previous synthetic data approaches. These results highlight the promise of integrating MRI physics into synthetic data generation for robust, generalisable stroke lesion segmentation. Code is available at https://github.com/liamchalcroft/qsynth

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4471_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/liamchalcroft/qsynth

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ChaLia_DomainAgnostic_MICCAI2025,
        author = { Chalcroft, Liam and Crinion, Jenny and Price, Cathy J. and Ashburner, John},
        title = { { Domain-Agnostic Stroke Lesion Segmentation Using Physics-Constrained Synthetic Data } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15963},
        month = {September},

}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors introduce two physics-constrained approaches to generate synthetic quantitative MRI (qMRI) data:

    • qATLAS: Trains a neural network to estimate qMRI parameter maps (PD, R1, R2*, MT) from standard MPRAGE images, enabling subsequent simulation of varied MRI sequences.
    • qSynth: Synthesizes qMRI maps directly from tissue labels using label-conditioned GMMs, with physical plausibility.

    The approaches were evaluated on four different datasets (ATLAS, ARC, PLORAS, and ISLES 2015) and compared against a baseline U-Net and SynthSeg. Results within and outside domain are convincing.

  • 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.

    Novel methodological contribution (physics-constrained synthetic generation and training, going beyond standard image-based data augmentation strategies) and in depth experiments and analysis (and clearly stated limitations).

  • 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 performance gains from combining synthetic and real data are inconsistent across datasets, making it difficult to determine when the inclusion of synthetic data is beneficial. It would good for the author to indicate why this is the case.

    • The paper’s organization and clarity could be improved. There are many tables (four in total), each with numerous rows, which affects readability. For example, not all contrast types are necessary for evaluating the method’s validity and could be omitted/merged to improve clarity. Similarly, the purpose of Fig. 2 is unclear — it’s difficult to interpret the generated scans, which limits its usefulness to the reader.

  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper presents a promising physics-constrained approach to improve the robustness of stroke lesion segmentation across heterogeneous domains. While further work is needed to address some limitations (including reducing the in-domain gap, although clearly described by the authors), this work makes a valuable contribution to the MICCAI community.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper introduces two innovative physics-constrained approaches for generating synthetic quantitative MRI (qMRI) data to address the challenge of stroke lesion segmentation across heterogeneous imaging protocols. The first method, qATLAS, utilizes a neural network to estimate qMRI maps from standard MPRAGE images, enabling realistic simulation of diverse MRI sequences. The second method, qSynth, directly synthesizes qMRI maps from tissue labels using label-conditioned Gaussian mixture models to ensure physical plausibility. Through extensive evaluation on multiple datasets, the authors demonstrate that both methods outperform baseline UNet segmentation, with qSynth notably surpassing previous synthetic data approaches, particularly in out-of-domain scenarios. By integrating MRI physics principles into the synthetic data generation process, these methods effectively bridge the gap between synthetic and clinical data, enabling more robust and generalizable stroke lesion segmentation in real-world clinical settings where acquisition protocols vary significantly.

  • 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.

    Novel Physics-Constrained Methodology: The authors introduce innovative approaches that leverage fundamental MRI physics principles to generate synthetic data with realistic tissue contrasts. This physics-informed framework represents a significant contribution to the field, with potential applications beyond stroke lesion segmentation to other medical imaging domains.

    Clear Problem Formulation: The authors present a well-articulated background and problem statement, effectively contextualizing the challenges of stroke lesion segmentation across heterogeneous MRI acquisition protocols. The manuscript establishes a compelling rationale for developing physics-constrained approaches to synthetic data generation.

    Methodological Rigor: The technical details of both proposed methods (qATLAS and qSynth) are meticulously described, providing comprehensive explanations of network architectures, parameter estimation techniques, and the integration of MRI physics principles. The training methodology is thoroughly documented, enhancing reproducibility of the work.

    Robust Experimental Design: The authors have implemented an extensive evaluation framework, testing their methods across multiple datasets (both public and proprietary) to demonstrate generalizability. The inclusion of both in-domain and out-of-domain testing scenarios strengthens the validity of their claims regarding improved robustness.

    Comprehensive Comparative Analysis: The manuscript features detailed comparisons against baseline methods and previous synthetic data approaches, with appropriate statistical analyses to substantiate performance improvements, particularly for the qSynth method in challenging cross-domain scenarios.

    Well-Structured Presentation: The organization of the manuscript is exemplary, with clear section delineation, informative figures illustrating the proposed methods, and detailed tables presenting quantitative results. The authors maintain a logical flow throughout the paper, culminating in a concise yet comprehensive conclusion that effectively summarizes the contributions and implications of the work.

  • 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 is free of major weaknesses. The authors have presented a methodologically sound study with well-designed experiments and clear presentation of results. The physics-constrained approaches for synthetic data generation represent a significant contribution to the field of medical image segmentation.

  • 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    The manuscript is free of major weaknesses. Congratulations on excellent work. I suggest a few minor improvements:

    Please clarify in the results section whether the provided DICE scores are specifically for lesion segmentation only, or if they represent an average across all tissue classes. Since the authors mention that qSynth provides additional tissue labels, it would enhance clarity to specify whether the reported metrics focus exclusively on lesion segmentation or incorporate multiple tissue classes.

    If possible, consider enlarging Figure 3 to improve readability and allow readers to better appreciate the qualitative differences between methods.

    To optimize manuscript space and enhance data presentation, the authors might consider consolidating ex. Tables 2 and 4 into a single comprehensive table.

    In the introduction’s first paragraph, I recommend citing the following relevant publications to strengthen the contextual framework:

    • https://ieeexplore.ieee.org/document/10230775
    • https://www.nature.com/articles/s43856-021-00062-8
  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This manuscript represents an exemplary contribution to the field of medical image analysis, specifically addressing the critical challenge of stroke lesion segmentation across heterogeneous MRI acquisition protocols. The authors present two novel physics-constrained methodologies (qATLAS and qSynth) that demonstrate significant improvements over existing approaches, particularly in challenging out-of-domain scenarios. The paper is exceptionally well-written and meticulously organized, with comprehensive technical details, rigorous experimental validation across multiple datasets, and clear presentation of results. The integration of MRI physics principles into synthetic data generation not only addresses an immediate clinical need but also establishes a framework that could be extended to other medical imaging applications. The authors’ thorough comparative analysis against existing methods, coupled with the demonstrated generalizability of their approach, provides compelling evidence of the work’s significance. Therefore, I confidently recommend accepting this paper independent of rebuttal, as it represents a substantial advancement in the domain of medical image segmentation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    Authors propose two MR-physics-based data synthesis approaches to enhance stroke lesion segmentation models’ generalizability across different datasets (with different image contrasts and imaging parameters). The two methods are: 1) qATLAS, which trains a U-Net to predict qMRI maps (including PD, R1, T2*, and MT) from MPRAGE images (trained with ground truth qMRI maps and simulated MPRAGE) and is applied to ATLAS dataset to generate predicted qMRI maps, and 2) qSynth, which directly generate synthetic qMRI maps from segmentation labels using Gaussian Mixture Models, with qMRI priors estimated from qATLAS and segmentation labels from ATLAS and OASIS. In both approaches, the synthetic qMRI maps are converted to synthetic MRI images (including FSE, GRE, FLAIR, and MPRAGE), which are used to train stroke lesion segmentation models. The segmentation models trained on real, different synthetic, or combined datasets are evaluated on both in-domain and three out-of-domain datasets. The two synthetic approaches are not as good as real data in the in-domain task but outperform the real data baseline in multiple out-of-domain tasks, with qSynth surpassing other methods.

  • 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 provides a good showcase on how MR physics-based data synthesis approaches can be implemented in practice for data augmentation in DL training. It provides two implementations to tackle the challenges of scarcity of qMRI data in such synthesis and did comparisons between the two. The attempt to estimate qMRI data from available weighted images, which can then be used for synthesis, is interesting, although qATLAS didn’t perform as well as qSynth in this work. The methods and the evaluation results can serve as a reference to other researchers in the field who are interested in investigating similar approaches.
    • Comprehensive evaluation on multiple datasets: As a work to enhance generalizability of DL models, the paper uses three out-of-domain datasets with multiple image contrasts for evaluation, including ARC (T1w, T2w; 229 subjects), PLORAS (106 T2w, 300 FLAIR), and ISLES2015 (T1w, T2w, FLAIR, DWI; 28 subjects).
    • Usage of multiple resources to create synthetic data: Authors take advantage of a bunch of public datasets and toolboxes to create the physics-based synthetic data, e.g., hMRI toolbox for ground truth qMRI generation, NiTorch for image simulation, ATLAS dataset for lesion labels, Multibrain SPM toolbox for healthy tissue maps, and the evaluation datasets. This information and the way they use them can be useful and inspiring to other researchers.
    • Extendibility: The MR physics-based data synthesis approaches for data augmentation can be useful in other disease, body parts, and applications.
  • 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.
    • Confounding factors in comparisons: The key difference between the two proposed approaches qATLAS and qSynth is their different synthetic strategies in creating qMRI maps. However, in the comparisons, there are more confounding factors. First, there is a noticeable difference in training dataset size: baseline and qATLAS segmentation models were trained on the ATLAS dataset (419/105/131 for train/val/test) while Synth and qSynth were trained on data created from OASIS-3 dataset (2579/100 for train/val). Second, the qATLAS model performs stoke lesion segmentation against background only while the qSynth model additionally predicts healthy tissue classes (GM, WM, GM/WM partial volume, and CSF). Third, the last sentence of Section 4 mentions that “The dataset’s skull-stripped images likely conferred an advantage to Synth/qSynth models compared to other methods”. These confounding factors make it unclear how different components contributed to the results.
  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    Thanks for the authors’ work and here are some thoughts:

    1. It would be helpful to briefly mention the key difference between Synth and qSynth approaches in the paper.
    2. qATLAS estimates [PD, R1, R2, MT] from a single T1-weighted MPRAGE image, where T1 (or R1) is the dominating information. Did qATLAS meet challenges in accurately predicting T2 or T2 information? (I understand that this can not be shown in the paper due to page limits.) The paper shows that qATLAS did well in T1w across the datasets but was inferior to Synth and qSynth in T2w and FLAIR (both more focusing on T2). I am wondering whether the accuracy difference between estimating T1 and T2 plays a role here.
    3. Before reading the results, I may expect qATLAS to have more realistic qMRI maps and corresponding synthetic images, e.g., with more realistic texture information, than qSynth. But the results showed that qSynth did better in many of the tasks. It would be interesting to investigate the effects of training dataset size in these experiments and there can be more discussions on why qSynth did better than qATLAS if the work is to be turned into a journal paper in the future. The ability to generate a large dataset by randomly combining segmentation labels and healthy tissue masks is also an advantage of qSynth.
  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While the idea of using qMRI data and MR physics to synthesize new data for training is not ‘super new’, I appreciate the authors’ efforts in implementing such idea in practice, with two different variants, and testing them out in the application of data augmentation and out-of-domain generalization. The paper is informative to others 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.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

N/A




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    First of all, the paper has been unanimously recognised to a good contribution by all reviewers. Highlights of the paper include a novel methodological approach (backed by image physics principles) a comprehensive and in depth experimental setup and analysis, an extensive use of libraries and coding resource (publicly available at review time if downloaded from the given link), its potential to other applications and a great presentation and an overall well-written manuscript. This is one the best papers on my stack and it is also one of the best contributions I have seen in my last few years as a reviewer. Therefore, I believe it deserves a provisional accept.

    Having said that, I would also like to point out that the reviewers have given some great suggestions to correct minor mistakes or minor weaknesses that could be addressed by the authors. I would advise the authors to take them into account to polish the current manuscript and also to consider giving a short explanation about the differences in training raised by reviewer #3 between the two approaches and give some thought for future work on why the gains of synthetic data are not universal among all datasets (as pointed out by reviewer #1).

    As an aside, I would like to remark that the anonymised url provided with the code can be glitchy when checking the website. Some of the files appear as “missing” online. However, this issue seems to only happen on the website and downloading the whole repository seems to fix it.



back to top