List of Papers Browse by Subject Areas Author List
Abstract
Magnetic resonance imaging (MRI) raw data (k-Space) contains valuable magnitude and phase information, yet clinical AI methods typically discard phase data. To address this, we introduce a magnitude-conditioned generative phase modeling framework that enables synthesis of complex-valued k-Space data from widely available magnitude-only MRI datasets. Pretraining with PhaseGen significantly improves k-Space skullstripping accuracy (from 40.1% to 80.1%) and enhances MRI reconstruction. This approach successfully bridges the gap between abundant magnitude-only datasets and information-rich complex-valued MRI data, enabling more accurate diagnostic tasks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4662_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TIO-IKIM/PhaseGen
Link to the Dataset(s)
N/A
BibTex
@InProceedings{RemMor_PhaseGen_MICCAI2026,
author = { Rempe, Moritz AND Hörst, Fabian AND Becker, Helmut AND Schlimbach, Marco AND Rotkopf, Lukas T. AND Kröninger, Kevin AND Kleesiek, Jens},
title = { { PhaseGen: A Diffusion-Based Approach for Complex-Valued MRI Data Generation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript introduces a magnitude-conditioned generative model that is capable of simulating or generating complex-valued k-Space data.
- 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.
1.overall, the manuscript is clearly presented and easy to follow. 2.the proposed magntude-guided phase generation model demonstrated the best performances for most experiments.
- 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.
1.There is not any simulated or generated phase images presented in the manuscript (only the PSNR / SSIM metrics showed), making the overall generation results doubtful. 2.In table 1, Naive [209k] demonstrated the best PSNR at 8x acceleration instead of the proposed PhaseGen, and an SSIM of <56% is also at a very low level. 3.No sequence details for the data involved were described, which is very important for this applicatio. 4.It is not clear whethter the output of the complex-valued network in the proposed model is the complex-valued data or phase images directly in the current version. please clarify. 5.The phase generation task presented in this manuscript is just a toy model instead of a realistic project in MRI reconstruction / acquisiitons, because one never acquire a “magntiude” image, only the “raw complex-valued kspace” data will be acquired in the forms of real + imaginary components instead of magnitude + phase. So it is just a click of the saving button in the scanner system to direct generate real phase data by just a fast fourier transform. 6.The authors are correct that the phase information is very important; however, this is not ture for all types of MRI images (acquired using different sequences), and unfortunately, I do not think the phase is important in fastMRI dataset (for most of them are acquired using FSE sequence). In addition, the downstream tasks are also too simple. Taking the most famous MRI phase-based sequence (Susceptibility Weighted Imaging, SWI) as an example, the most important or foundamental informaiton is hidden in a background part, whos signal / intensity is usually much higher than the true information, and thus formulating the well known phase wraps.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, the phase generation task presented in this manuscript is just a toy model instead of a realistic project in MRI reconstruction / acquisiitons for majorly two reasons: 1.MRI scanners never acquire a “magnitude” image; only the “raw complex-valued kspace” data will be acquired in the form of real + imaginary components instead of magnitude + phase. So it is just a click of the saving-data button in the scanner system to directly generate real phase data by just a fast Fourier transform. 2.Indeed, the phase information is not always very important for all types of MRI images (acquired using different sequences), and that is why there are only magnitude images in most previous public datasets. A realistic MRI phase-based application is SWI or more recent QSM (quantitative susceptibility mapping), and in these two sequences, the “phase” data (different from magnitude data) is much more complicated because a background signal should also be formulated and considered (this is because of the non-locality nature of phase components and partly the reason for phase wraps). These two points make this manuscript a toy-model-level work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors argued, “Regarding the absence of qualitative phase visualizations: raw MRI phase maps are inherently non-interpretable to human observers, as their appearance is dominated by background field contributions unrelated to tissue structure”. However, this is not always the case, especially when multi-echo GRE sequences are used, and we can observe the obscure fine structures in tissues. The case is also true for FSE data, because they reflect the information of B0 field in the scanner. So, it is very important to “see” the phase images because they are meaningful in many important applications.
I agree that the complex-valued data should be beneficial for improving the results of the classification and cardiac segmentation tasks, as reported in recent works mentioned by the authors, but it does not mean that the “synthesized” phase data will also improve the performance of models on real-world applications. Indeed, estimating phase data from magnitude should be a very ill-conditioned problem, because they represent information from different dimensions/aspects, respectively. It is nearly impossible to complete this task without strong assumptions or priors. For example, knowing a complex number z’s magnitude is 1 only identifies a unit circle in complex space, instead of the exact phase. If this task is feasible, then the authors should discuss the reason and formulate the process.
As for the downstream brain extraction task shown in this work, the improvements in numeric metrics like DSC do not mean that you have got better brain extraction results, which can only be identified by visualizing the original brain images and the brain extraction/segmentation maps. Besides, the BET task itself is too simple; the authors should at least conduct a “realistic” segmentation work to demonstrate the effectiveness of their model, although in my opinion, the role of the model is not learning the “phase”, it is more like a high-level latent information beneficial for brain segmentation.
It will be better to conduct background removal to remove the background field from the synthesized phases using existing algorithms to see if they really synthesized useful information from the magnitude data, which will be a much more meaningful and persuasive downstream experiment compared to the BET experiment.
Review #2
- Please describe the contribution of the paper
The paper introduced a magnitude conditioned generative phase modeling framework that enables synthesis of complex-valued k-Space data from widely available magnitude-only MRI datasets. Using the proposed method to generate complex training data and train for downstream tasks can significantly improves k-Space skullstripping accuracy and enhances MRI reconstruction.
- 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 targets a very practical and important issue in MRI deep learning community: The lack of complex data for training and the performance degradation of magnitude-data-trained network on real-world application. It provides a novel way to generate complex dataset using diffusion model and shows its benefits on downstream tasks.
- 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.
1.Lack of qualitative results
The paper does not include example phase images generated by the diffusion model or comparisons with ground truth. Such visualizations are important for assessing reconstruction quality in medical imaging and should be included.
2.Writing and organization
The writing can be improved for clarity. In Material and Methods section, I would suggest use separate sections to introduce methods/dataset for proposed phase generation method and downstream validation methods, which can benefit readers. In Results, avoid repeating table values in the text and instead focus on summarizing key findings and insights.
- 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.
- 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.
(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?
The paper targets a very practical and important issue in MRI deep learning community: The lack of complex data for training and the performance degradation of magnitude-data-trained network on real-world application. It provides a novel way to generate complex dataset using diffusion model and shows its benefits on downstream tasks.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes PhaseGen, a magnitude-conditioned diffusion framework for generating synthetic phase and thus synthetic complex-valued MRI data from magnitude-only datasets. The generated complex data are then used to support downstream k-space learning tasks such as MRI reconstruction and skull stripping, aiming to bridge the gap between magnitude-only datasets and raw-data-based learning.
- 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.
(1) The paper aims to leverage large-scale magnitude-only MRI datasets for complex-valued raw-data learning, which is an underexplored research topic. (2) The proposed magnitude-conditioned phase generation framework is conceptually clear, providing a practical way to synthesize complex-valued data. (3) PhaseGen is evaluated on downstream tasks (reconstruction and skull stripping), which is more meaningful than evaluating generation quality alone. Besides, it shows consistent improvements over naive baselines, indicating that the generated phase carries useful information.
- 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.
(1) A critical control experiment is missing: the paper does not compare “partial real data only” versus “partial real data and synthetic data”, making it difficult to validate the claimed data efficiency. (2) The downstream evaluation is limited, focusing on single-coil reconstruction and a relatively specialized skull stripping task, which restricts the generality of the conclusions. The skull stripping evaluation relies on pseudo ground truth (STAPLE fusion) instead of manual annotations, which weakens the reliability. (3) There is limited analysis of the realism and failure modes of the generated phase across different imaging conditions. (4) It would be valuable to clearly discuss recent physics-informed synthetic data approaches (e.g., PISF on MedIA 2025, doi: 10.1016/j.media.2025.103616), which are highly relevant to synthetic data generation for real-world MRI reconstruction. Besides, expanding validation to more realistic settings (e.g., multi-coil reconstruction and broader downstream tasks) and providing more analysis on the realism and limitations of the generated phase would strengthen the work.
- 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.
- 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?
This PhaseGen framework offers a reasonable and technically sound solution for “how to leverage magnitude-only MRI datasets for complex-valued raw-data learning”. The method demonstrates consistent improvements over naive baselines and shows potential as a data augmentation strategy for k-space learning. However, the experimental validation is not yet fully sufficient to support the broader claims, while the evaluation is also somewhat limited in scope, and the novelty is moderate. Despite these limitations, I believe the paper could facilitate further research in this direction.
- 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.
N/A
Author Feedback
We thank all reviewers for their constructive feedback and address the major concerns below.
MR / R1 + R2: Phase Visualization & Network Output (R1-W1, R2-W1, R1-W4) PhaseGen outputs full complex-valued signal (real + imaginary), not phase images alone. Regarding the absence of qualitative phase visualizations: raw MRI phase maps are inherently non-interpretable to human observers, as their appearance is dominated by background field contributions unrelated to tissue structure. We therefore prioritized downstream functional validation as a more meaningful and direct quality measure. The downstream results directly quantify phase quality: PhaseGen achieves skull-stripping DSC = 80.1% vs. 41.1% for naive generation (same data volume) and 40.1% for no-phase training. This is a ~100% relative improvement demonstrating anatomically coherent phase. Reconstruction results in Tab. 1 confirm this across all metrics, model sizes, and undersampling rates.
MR / R1: “Toy Model” & Clinical Relevance (R1-W5 & W6) The reviewer argues phase is recoverable via FFT since scanners acquire raw complex k-Space. This is correct in principle but inapplicable to our problem. Our method targets existing magnitude-only public datasets (1.68M subjects [7]) where raw k-Space was never stored. Even where scanner access exists, curating a labeled, IRB-approved, quality-filtered raw k-Space dataset is highly time-intensive. The clinical relevance is concrete: FSE/TSE, FLAIR, and TIRM sequences dominate these archives. Recent work shows that complex-valued raw data improves downstream task stability at high undersampling rates, with demonstrated gains in prostate classification and cardiac segmentation over magnitude-only baselines [5, 19]. PhaseGen enables these clinically validated approaches to scale to magnitude-only data without new acquisitions, labeling or IRB re-approvals. These are substantial real-world barriers. For SWI/QSM, we agree phase synthesis is more complex. PhaseGen does not target these.
R1: Worse PSNR in Tab. 1 (R1-W2) We respectfully note a factual inaccuracy: at 8x undersampling, PhaseGen 209k achieves PSNR = 20.61 dB vs. Naive 209k = 20.41 dB. PhaseGen is higher. The low SSIM at 8x is expected for such high acceleration in a single-coil setting, yet PhaseGen still provides a measurable improvement over non-original baseline.
R1: Missing Sequence Details (R1-W3) Training sequences: TSE, TSE SPAIR, FL2D, TIRM, FLAIR; slice thickness 2–8 mm.
R1: No code available The anonymized repository is linked on p.2 (end of Sec.1), acknowledged by R2 and R3. R2-W2: Writing and organization improvements We thank the reviewer for his positive feedback and specific suggestions to improve our manuscript.
R3: Missing Partial Real Data Only Control (R3-W1) This is a valid point. We acknowledge that Table 2 does not include a “partial real data only” baseline. We argue that the skull stripping experiment (Fig. 2) provides the relevant evidence: naive generation and PhaseGen use identical data volumes, yet DSC differs by 39pp. The benefit is attributable to phase quality, not data volume.
R3: Single-Coil & Pseudo Ground Truth (R3-W2) We chose Single-Coil data to isolate phase synthesis quality from coil sensitivity confounds. We agree that human annotated data would be the gold-standard for the experiments. The STAPLE consensus of three state-of-the-art skull stripping algorithms is an established silver-standard proxy when manual annotations are not available.
R3: Phase realism, failure modes & related work (R3-W3 & W4) Observed failure modes include phase artifacts at high-contrast tissue boundaries and reduced accuracy for sequences outside the training distribution. Regarding PISF (MedIA 2025): Thank you for mentioning this very interesting work. But unlike physics-informed approaches that simulate acquisition, PhaseGen generates phase conditioned on magnitude from existing datasets without scanner modeling.
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.
The reviews are mixed, with scores of 2, 5, and 4.The paper tackles an interesting and practical problem: using magnitude-only MRI datasets to synthesize complex-valued data for downstream k-space learning. The proposed magnitude-conditioned phase generation framework is conceptually clear, and the downstream reconstruction and skull-stripping experiments suggest that the generated phase may provide useful signal beyond naive baselines.
However, several important concerns need to be addressed before a final decision. The rebuttal should clarify the physical and clinical relevance of the phase-generation task, including which MRI sequences benefit from synthetic phase and how the method relates to realistic raw complex-valued acquisition. The authors should also provide qualitative examples of generated phase images, analyze failure modes, and clarify whether the network outputs phase images or complex-valued data.
- 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.
I recommend acceptance. The paper addresses a practical and underexplored problem: how to leverage large magnitude-only MRI datasets for complex-valued/k-space learning. The proposed magnitude-conditioned generation of complex-valued data is conceptually clear, and the downstream reconstruction and skull-stripping experiments suggest that the generated complex signal provides useful information beyond naive phase generation or no-phase baselines.
The rebuttal clarified several important points, including that PhaseGen outputs complex-valued real/imaginary data rather than phase maps alone, the intended scope of the work as augmentation for existing magnitude-only datasets, the MRI sequence details, and the limitation that the method is not aimed at SWI/QSM-like phase applications. Reviewers #2 and #3 support acceptance after rebuttal. Reviewer #1 remains skeptical about physical realism and phase interpretability, which is a valid concern, but I view the method as a useful data-generation/augmentation contribution rather than a claim of uniquely recovering true physical phase.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The evidence remains mixed after rebuttal on whether the generated phase is physically and clinically realistic across imaging conditions. The lack of qualitative phase or error analysis, limited failure-mode discussion, single-coil reconstruction setting, specialized skull-stripping task, and use of STAPLE pseudo ground truth leave important questions open.
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.
This paper maintains mixed reviews after rebuttal. It does appear, after the rebuttal, that there is a misunderstanding regarding the purpose of the paper, which the authors argue is to generate complex k-space data from magnitude images. This is a highly ill-posed problem, but it is suitable for machine learning-based techniques such as diffusion models. This was clear to me from the rebuttal. However, I caution the authors on maintaining that the lack of qualitative phase data is reasonable. Qualitative data, regardless of interpretability, is vital for synthesis-based works, and I encourage the authors to include it in the final paper. That being said, I do believe that the authors have adequately addressed the reviewers concerns and I recommend acceptance of this paper.