List of Papers Browse by Subject Areas Author List
Abstract
Deep learning–based medical image segmentation is increasingly used to support clinical diagnosis and develop new treatment strategies. However, model performance remains limited by the scarcity of high-quality annotated data and insufficient generalization across imaging protocols. This limitation is particularly evident in MRI and CT, where models are typically trained on a single acquisition sequence and exhibit reduced robustness when applied to unseen sequences or contrasts. Although data augmentation is widely used to improve general robustness on medical images, its impact on cross-modality generalization has not been quantitatively explored.
In this work, we study a targeted set of data augmentation techniques designed to improve cross-modality transfer. We train three spine segmentation models, each on a single-modality/sequence dataset, and evaluate them across seven out-of-distribution datasets (spanning CT and MRI), reflecting a realistic single-sequence training and multi-sequence/contrast/modality deployment scenario. Our results demonstrate substantial performance gains on unseen domains (average Dice gain of 155%) while preserving in-domain accuracy (average Dice decrease of 0.008%), including effective transfer between CT and MRI.
To mitigate the computational cost typically associated with strong data augmentation, we implement GPU-optimized augmentations that maintain, and even improve, training efficiency by approximately 10%. We release our approach as an open-source toolbox, enabling seamless integration into commonly used frameworks such as nnUNet and MONAI. These augmentations significantly enhance robustness to heterogeneous clinical imaging scenarios without compromising training speed.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1360_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/neuropoly/AugLab
Link to the Dataset(s)
N/A
BibTex
@InProceedings{MolNat_One_MICCAI2026,
author = { Molinier, Nathan AND Möller, Hendrik AND Dagonneau, Thomas AND Curto-Vilalta, Anna AND Graf, Robert AND Atad, Matan AND Rueckert, Daniel AND Kirschke, Jan S. AND Cohen-Adad, Julien},
title = { { One Sequence to Segment Them All: Efficient Data Augmentation for CT and MRI Cross-Domain 3D Spine Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors of this paper develop an open-source toolbox for GPU-optimized data augmentations that are designed to improve cross-modality transfer. They train three separate models for spine segmentation and perform an evaluation on seven out-of-distribution datasets. They demonstrate that their method has substantial improvement for unseen domains, and minimal reduction in Dice scores on in-domain data. The authors also conducted extensive ablation studies to assess the effect of each augmentation.
- 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 authors perform a significant analysis of their model. This includes comparing models trained on three different datasets, and evaluation on the baseline, nnUNetTrainerDA5, and their proposed method (Table 3). They additionally compare the vertebral, IVD, and canal segmentations using models trained on these three datasets. The reviewer appreciates that the evaluation was performed per scan sequence type.
- The authors perform a significant number of ablation studies using their approach (Table 5). They add in each ablation separately to the baseline model to test the effect, and do so for all the models trained on SG in-phase and SG CT.
- The authors provided clear qualitative examples comparing the models trained on the three datasets for the nnUNet baseline vs their method.
- The manuscript includes a statistical analysis to compare methods.
- The authors provide a good discussion.
- They make their code open source.
- 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.
- Though the authors commented on the field of domain randomization and other synthetic data generation methods, the manuscript would be improved with a quantitative comparison.
- The authors include in their limitations that the generalizability of their method is unknown to other anatomical regions. However, more information should be included about additional problems they may encounter with other datasets.
- 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
- It would be useful to include the standard deviation in Tables 3,4, and 5.
- 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?
Though the authors use novel transforms, perform significant evaluation of their models, and use multiple datasets, the paper could be slightly improved by a comparison to other data augmentation and domain randomization approaches.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
Thank you for the rebuttal. However, I believe that due to evaluations of one task, a lack of comparison to other synthetic data-based methods/diffusion methods, the manuscript is not ready for acceptance at this moment.
Review #2
- Please describe the contribution of the paper
The paper introduces a training-free, model-agnostic data augmentation (DA) strategy that is shown to improve generalization on a 3D spine segmentation task. The proposed strategy involves the extension of classic approaches with a list of data augmentation transforms that mostly operate on the intensities or texture of the image (i.e. no additional morphological or geometric deformations). The inevitable additional computation cost is mitigated by an optimized re-implementation of the full DA pipeline, which is shown to significantly reduce the application runtime.
- 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 manuscript presents several notable strengths. Its focus on usability is commendable, as the proposed approach is model-agnostic and benefits from an optimized GPU implementation. The optimized GPU implementaiton is a very valuable contribution in its own right. The comparison with the nnUNet baseline is rigorous, spanning multiple modalities and supported by statistical significance testing. The work is further strengthened by extensive ablation studies, and the authors provide a balanced discussion of both results and 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 manuscript also presents several weaknesses. On the major side, the experimental validation is limited to a single type of task, and the comparison with the state of the art is insufficient — notably absent are comparisons with diffusion-based approaches for cross-modal generation [1] and SynthSeg [2]. Furthermore, the ablation studies suggest that RedistributeSeg alone accounts for a large share of the generalization improvement in most settings, which weakens the case for the full proposed transformation pipeline. Since RedistributeSeg is not a contribution of this work, this observation somewhat diminishes the overall impact of the paper. On the minor side, the evaluation on canals raises a methodological concern: given that these structures may not be visible in the CT image, it is unclear whether such evaluation is meaningful, and how the ground truth was defined is not discussed. Finally, the authors do not adequately address a noteworthy ablation result, namely the surprisingly poor performance of single transformations on the SG-CT to SG-water generalization task, which stands in stark contrast to the strong boost observed when all new transformations are compounded. The nature of the interaction between transformations that could explain this behaviour, and why it appears to be specific to this particular setting, warrants further discussion.
[1] Ma, Ji, Jinjin Chen, and Aoxiang Liang. “CT-to-MRI translation of medical volume data based on an enhanced diffusion model.” Scientific Reports (2026). [2] Billot, Benjamin, et al. “SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining.” Medical image analysis 86 (2023): 102789.
- 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
I thank the authors for their effort in producing a reusable and reproducible scientific contribution. As a minor remark, in Table 3 the brown color is indicated to denote in-domain settings; however, it is unclear why MM CT is not highlighted in brown for the SG CT training example, as both are CT domain.
- 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 paper is well written, with a clear objective and a well-defined scientific question. However, the lack of evaluation across a wider range of tasks substantially limits the generality of the conclusions and reduces the overall impact of the work. Moreover, given that RedistributeSeg consistently dominates the generalization gains across most experimental settings, the paper reads less as a demonstration of the utility of the full proposed data augmentation pipeline, and more as an assessment of the effectiveness of RedistributeSeg alone.
- 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 concern about the impact and relevance for the MICCAI community has been significantly reduced by the meta review 1
- the concern about the utility of the full pipeline, given that RedistributeSeg often provides the major contribution, has been addressed by the authors’ rebuttal. I agree with them that discovering this fact is interesting in its own right and would have been impossible without their extensive ablation studies. Their comment about “there is no single best augmentation for MRI or CT” is also valid.
- the lack of comparison with SynthSeg remains a weakness of the paper, although overall the rebuttal is sufficient (in my opinion) to tip the scales towards acceptance. The authors mention SynthSeg’s higher computational and setup requirements: I am not aware of any particularly strong requirements in this sense (especially compared to other state of the art methods), but if the authors feel strongly about this they could introduce a sentence justifying the lack of comparison.
- as an additional comment that does not influence my suggestion for a decision on the rebuttal: please include a few sentences explaining how the annotations for canals on CT were defined by the radiologists, given that canals are invisible on CT. Did they use some proxy anatomical landmarks? Or just their own experience? Or some other strategy?
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is demonstrating that well‑designed, appearance‑based data augmentation is sufficient to achieve strong cross‑domain generalization for 3D spine segmentation, even when training is limited to a single CT or MRI sequence. The proposed augmentation pipeline is implemented efficiently on the GPU within the nnUNet framework. While individual augmentations lead to only modest performance improvements, the results show that the largest gains arise from combining all augmentations together, with the specific order of their application having little impact on performance.
- 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.
A major strength of this paper is the thorough and well‑designed ablation study, which systematically evaluates a set of carefully chosen data augmentation techniques targeting contrast and texture appearance variations. These augmentations are explicitly motivated by the challenges posed by multi‑contrast, multi‑parametric MRI as well as cross‑modal differences between MRI and CT, and the study clearly demonstrates how they contribute to improved robustness.
- 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 does not introduce a fundamentally novel methodology. The use of data augmentation to improve contrast‑agnostic segmentation performance has been explored in prior work; for example, Bédard et al. (2025) demonstrate robust spinal cord segmentation across MRI contrasts using aggressive augmentation and contrast‑aware training strategies. As a result, the contribution of this work is primarily empirical rather than conceptual. Another limitation is that the evaluation is restricted to a single anatomical structure and task, namely semantic spine segmentation. Given the paper’s emphasis on carefully designed augmentations, it would strengthen the contribution to provide guidance on how these augmentation choices generalize to other anatomies or tasks, or to discuss strategies for automating augmentation selection rather than relying on task‑specific manual design.
Ref : Bédard, S., Naga Karthik, E., Tsagkas, C., Pravatà, E., Granziera, C., Smith, A., Weber II, K. A., & Cohen-Adad, J. (2025). Towards contrast-agnostic soft segmentation of the spinal cord. Medical Image Analysis, 101, 103473.https://doi.org/10.1016/j.media.2025.103473
- 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.
(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?
I recommended a Weak Reject primarily because, while the paper presents a strong and carefully executed empirical study, its core contributions are incremental rather than clearly novel. The use of appearance‑focused data augmentation to achieve contrast‑agnostic or cross‑modality segmentation has been demonstrated in prior work, and the paper does not introduce a new formulation, learning paradigm, or theoretical insight beyond this existing direction. Another factor influencing my score is that the evaluation is limited to a single anatomical structure and task, making it difficult to assess how broadly the proposed augmentation strategy generalizes. I would not strongly oppose acceptance, particularly if the rebuttal can better clarify the paper’s positioning relative to prior work or better articulate the generality and broader impact of the proposed approach.
- 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
Author Feedback
The reviewers note the lack of comparisons with SynthSeg and diffusion-based cross-modal generation methods, and that the contribution appears empirical rather than novel, citing a paper by Bédard et al. The validation is limited to a single anatomy, leaving the method’s generalizability unclear. They request a broader discussion of how the augmentation design may transfer to other anatomies.
The primary goal of this work was to perform a detailed empirical study across three datasets, seven sequences, and three segmentation structures to evaluate whether a simple plug-and-play augmentation strategy could generalize without requiring complex pipelines or fine-tuning. We prioritized depth of analysis over broader comparisons. We agree that extending the study to non-spinal tasks would strengthen it; However, evaluating our pipeline on three distinct structures, each exhibiting different behaviors with respect to contrast and modality, represented a reasonable trade-off for assessing how well the performance generalizes across different anatomical structures, given the 8-page limit. Future work will investigate additional anatomical structures. We acknowledge that including additional comparative methods, such as SynthSeg, would further strengthen the study. However, given the trade-off imposed by our exhaustive benchmark, we chose to prioritize the inclusion of multiple plug-and-play data augmentation over more complex methods. When designing the study, SynthSeg was not considered a valid competitor due to its higher computational and setup requirements.
Regarding comparisons with Bédard et al.: They simultaneously train on six image contrasts, whereas our problem statement is to demonstrate improved cross-domain transferability when training on a single contrast. In addition, they did not investigate appearance-altering augmentations for improving cross-domain transferability. We will add these details to the literature review section.
The reviewers correctly noted that the RedistributeSeg augmentation alone contributes most to the domain-transfer capabilities. They argue that this weakens the motivation for using the proposed combination of augmentations, especially since RedistributeSeg is not the main contribution of this work.
Our ablation study revealed the strong contribution of RedistributeSeg, which would likely have remained unnoticed without such extensive analysis. We agree that this aspect is mostly true across contrasts. However, the example of training on SG-CT and evaluation on SG-water, also mentioned by the reviewers, provides a counterexample to this statement. Table 5 demonstrates that while RedistributeSeg is the best-performing across most sequences, this is not the case for SG in-phase and SG water when trained on SG CT. This shows that there is no single best augmentation for MRI or CT, underscoring the importance of such extensive ablation. In addition, our study shows that our proposed combination often significantly outperforms RedistributeSeg alone. The discussion will be extended to better explain our results.
The reviewers questioned the ability to evaluate structures such as the spinal canal on CT images, as they are not visible. They also requested the standard deviation in Tables 3, 4, and 5 and noted that labeling a CT as out-of-domain for another CT dataset is confusing. We agree and mention in the discussion that the CT canal annotations are probabilistic. The fact that the model trained on CT with our augmentation achieves great values on MRI still validates our approach. Moreover, Figure 3 shows that differences in CT predictions compared to the default nnunet are due to real performance issues: the default nnunet does not segment the canal. We will clarify this in the revised manuscript.
We removed the standard deviations to improve readability, but we agree with the reviewers and will add them back in. We will also clarify the color scheme regarding the definition of “in-domain”.
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 introduces a GPU-optimized, model-agnostic data augmentation toolbox aimed at improving cross-modality generalization for 3D spine segmentation. The approach focuses on appearance-based transformations and demonstrates improved robustness on multiple out-of-distribution datasets, while maintaining strong in-domain performance. The authors also provide extensive ablation studies and release their implementation as open source.
All reviewers agree that the paper addresses a relevant and practical problem, and highlight the thorough experimental evaluation as a key strength. In particular, the study includes multiple training settings, evaluation across several datasets and modalities, detailed ablations, and statistical analysis. The practical utility of the method—being training-free, easily integrable, and GPU-efficient—is also appreciated.
However, there is a general consensus that the methodological novelty is limited. I personally do not see this as a limitation because I see this paper as an empirical study, which are welcome at MICCAI. I think that the combination of the GPU-based nnUNet implementation and the thorough evaluation make this paper a good contribution.
Nevertheless, I would like the authors to address several points in their rebuttal. First, I would like them to better position themselves with the literature and to comment on potentially missing baselines: SynthSeg (which is very close in concept and implementation), other augmentation strategies for spinal cord, and cross-modal generation. Additionally, one reviewer notes that a significant portion of the observed gains may be attributed to a single component (RedistributeSeg), which weakens the case for the full pipeline. I would also like the authors to comment on the generalisation to other anatomies and tasks, although I think that the current setup is enough for MICCAI, especially given the depth of the conducted experiments. Finally, small concerns are also to be addressed, such as missing standard deviations, result clarifications, and dataset constructions.
Overall, the paper presents a solid and well-executed empirical study with clear practical value. However, concerns about positioning, clarification of contributions, and discussion of generality, should be addressed in rebuttal.
- 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 authors successfully answered a number of concerns raised in the initial reviews about: 1) including a discussion of the RedistributeSeg augmentation step, 2) clarifying the difference with existing existing works (Bédard et al.), 3) clarifying spinal cord annotations, 4) adding the missing standard deviations. As said earlier, I also think that empirical studies are welcome at MICCAI, especially when they address widespread subjects such as data augmentation (like in this study). Moreover, given the depth of the conducted analysis, I think that the author’s position of keeping the extension to other organs for future work is reasonable.
However, not including a SynthSeg-like method for cross-modal segmentation is a clear flaw for this paper. As pointed out by reviewer 2, it is not true that SynthSeg requires special computational resources, and can be trained on a normal GPU. This flaw is amplified by the fact that abdominal SynthSeg methods [a] have already been trained and are clearly missing here.
Overall, this paper is borderline accept/reject. However, I lean towards acceptance given the broad interest for a thorough empirical study of data augmentation, which goes way beyond the domain randomisation (SynthSeg) community.
[a] Benchmarking of deep learning methods for generic MRI multi-organ abdominal segmentation, Krishnaswamy et al, Journal of Medical imaging, 2025.
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.
R1 changed from Weak Accept to Reject, R2 changed from Weak Reject to Accept, while R3 did not change their initial recommendation of Weak Reject. Although the paper is promising, with two reviewers ultimately recommending Reject, I would also recommend rejection.
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.
My decision relies in large part to the original Area Chair comments, although it takes into account the reviewer’s concerns and the rebuttal answers to them.
For starters, I agree with the original Area Chair, that the framing for the paper is key and as provided by the authors in their rebuttal, I would consider this work an empirical study and not a methodological contribution. In that sense, methodological novelty should not be the main reason for rejection or acceptance.
Based on the empirical study, I also agree that contextualisation is key to make a decision. The authors provided with their rebuttal a reason for why domain randomisation techniques were not included. While I am sympathetic with not including them due to the additional training cost (not viable during rebuttal), I am not entirely sure I can agree with the methods being that different, considering RedistributeSeg as one of the augmentations and the combination of all augmentations if applied to a case. To clarify, I think part of the reasoning is overall sound and adding a few sentences to highlight that would be beneficial, but I would emphasise that there are similarities and that comparison should be performed in a future work (perhaps a journal publication).
I am also not entirely convinced by the argument about a couple of specific cases where RedistributeSeg is not the best augmentation. From the whole ablation table, these are only 2 experiments and in both cases, RedistributeSeg is the second best augmentation not that far behind. Nonetheless, I do agree with the authors that the thorough ablation test is commendable and leads to a deeper and richer analysis.
Apart from that, I believe the authors did a good job of addressing all the minor and major concerns raised by the reviewers and original Area Chair.