List of Papers Browse by Subject Areas Author List
Abstract
Segmentation of vestibular schwannoma and cochlea from T2 MRI is clinically important yet annotation-intensive. Domain adaptation (DA) has been widely adopted to bridge the gap between labeled contrast-enhanced T1 and unlabeled T2 datasets. While existing methods focus on cross-domain alignment, intra-domain variability within the target domain remains largely overlooked. Images from the same domain may vary substantially due to different scanners, field strengths, and acquisition protocols. Ignoring this variability produces homogeneous synthetic images that limit the generalizability of downstream segmentation models. To address this, we propose IntraStyler, a 3D unpaired image translation method that automatically discovers fine-grained intra-domain styles without any predefined sub-domains, and synthesizes diverse target domain images using per-image style references. To this end, we design a 3D style encoder trained with a novel contrastive learning objective to extract style-only embeddings disentangled from anatomy. IntraStyler is built upon the 1st place CrossMoDA challenge solution and further advances it, generating more diverse synthetic data and achieving more reliable downstream segmentation. Code is available at https://github.com/MedICL-VU/IntraStyler.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2804_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/MedICL-VU/IntraStyler
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuHan_IntraStyler_MICCAI2026,
author = { Liu, Han AND Fan, Yubo AND Li, Hao AND Hu, Dewei AND Moyer, Daniel AND Xu, Zhoubing AND Dawant, Benoit AND Oguz, Ipek},
title = { { IntraStyler: Intra-Domain Style Synthesis for Cross-Modality MRI Domain Adaptation } },
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 paper proposes to address the lack of fine grained domain discrimination in cross modality translation by leveraging a contrastive learning mechanism based on synthetic intensity transforms of the target domain (e. g. random contrast change, gaussian smoothing. . . ). This allows to learn variations of appearance through contrastive training. The style embedding is then handled by some form of adaptive instance normalization. The approach also allows for smoothly interpolating between two reference styles using SLERP. Authors apply their method to the CrossMoDA segmentation challenge setting by integrating the style encoder to the winning solution of the CrossMoDA challenge [16].
- 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 definition of source and target domains in I2I is often vague and embeds high heterogeneity. it is thus very sound to try to tackle domain variations in an unsupervised fashion. CrossMoDA is an appropriate setting as the target domain is heterogeneous in style. Method is very well described and sound and results clearly show the ability of the approach to generate diverse styles while preserving the structural fidelity of the original T1 image. Quantitative results are also very good. It is overall a very well written paper of outstanding quality, that could be applied in a variety of other image to image translation setting.
- 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 contrastive learning mechanism relies on synthetic data augmentation. The selection, relevance and relative importance of these have not been explicitly discussed by the authors.
- Applied synthetic transformations are global to the image. I suspect this explains that, From Fig. 3, it seems that the applied style is global, i. e. the VS appearance in the synthetic image (to the right) does not follow the VS appearance in the reference #2 image, i. e. the VS is hypersignal in the reference, but hyposignal in the target, as the synthetic transformation seemed to have applied a global darkening of the style. The variability in target appearances in crossMoDA has been previously discussed in another top performing method Salle et al. TBME 2024.The authors propose a more local alteration by only changing tumor appearance. Could such more local data augmentation solve the suspected problem of global contrast change ? I think that point could be mentioned as future evolution of the work or simply discussed.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
It is a very sound methodology that addressses the crucial challenge of ill defined target domains in cross modality translation. Code is available and I am confident the community can largely benefit from such a contribution.
- 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 #2
- Please describe the contribution of the paper
This paper proposes IntraStyler, a 3D unpaired image translation framework for cross-modality MRI domain adaptation targeting vestibular schwannoma and cochlea segmentation. The core contribution is a sub-domain-free style diversification mechanism: a 3D style encoder trained with a contrastive learning objective extracts per-image style embeddings disentangled from anatomical content, enabling style-conditioned synthesis without predefined sub-domain labels. The style encoder is integrated into the CrossMoDA challenge-winning backbone via dynamic instance normalization layers and jointly optimized end-to-end. Style interpolation via SLERP further enables smooth blending between reference styles. Downstream segmentation is evaluated on the CrossMoDA benchmark against three synthesis strategy variants.
- 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 observation that institute-level sub-domain labels are too coarse to capture genuine intra-domain variability is well-supported by Figure 1 and the experimental design. The problem of sub-domain label dependency is a real bottleneck for clinical deployment, and the proposed solution is principled. 2.Defining positives as same-image crops with different anatomy, and negatives as intensity-perturbed versions of the positive, is a clean and annotation-free strategy for learning style-invariant-to-anatomy embeddings. The fact that the encoder trained on synthetic perturbations generalizes to discriminate real MRI scanner differences (Fig. 4) is a notable and non-obvious finding.
- 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.IntraStyler is explicitly built upon the CrossMoDA challenge-winning method and the primary modification is the addition of a style encoder with contrastive training and DIN injection. The core synthesis backbone, discriminator, PatchNCE loss, and segmentation pipeline are all inherited. While the results are strong, the methodological novelty relative to prior style-conditioned synthesis work (e.g., StarGANv2, AdaIN) is incremental, and the paper’s framing as the “first 3D unpaired image translation method for sub-domain-free style synthesis” requires more careful qualification given the breadth of existing style transfer literature. 2.All experiments are conducted on a single benchmark (CrossMoDA) involving one specific anatomy (posterior fossa), one translation direction (ceT1→T2), and one imaging modality pair. The paper acknowledges that intensity perturbation functions may require modality-specific tuning for CT or PET, but provides no evidence of transferability beyond brain MRI. The generality claim in the conclusion is therefore unsupported. 3.At inference time, IntraStyler uses K-means clustering with silhouette analysis to select 7 representative reference images. The number of selected references (7 vs. 3 for sub-domain methods) is not controlled for in the comparison. It is unclear whether the performance gain comes from the quality of the style encoder or simply from using more diverse references. An experiment fixing the number of references or sampling randomly would help disentangle these factors.
- 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.
- 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?
IntraStyler addresses a well-motivated limitation of existing sub-domain-based style synthesis methods and demonstrates clear, statistically validated improvements on a challenging benchmark. The contrastive learning formulation is elegant, the experimental design is controlled, and the failure case reduction is clinically meaningful. The primary concerns are the limited scope of evaluation (single benchmark, single anatomy, single modality pair), the uncontrolled number of reference images in comparison, and the incremental nature of the architectural contribution relative to the baseline.
- 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
The paper proposes a 3D unpaired image translation framework for cross-modality MRI domain adaptation that models intra-domain variability in the target domain without relying on predefined sub-domain labels. The main contribution is a per-image style conditioning strategy in which target images serve as an implicit style reference, enabling the synthesis of diverse target-style images from labeled source data. Central to their approach is a contrastively trained style encoder designed to extract style representations that should be disentangled from anatomical content. The contrastive learning scheme operates on patches sampled from target domain images, where positives are different spatial crops from the same image and negatives are obtained by applying intensity perturbations to these crops. The learned style embeddings are subsequently clustered to identify representative styles that are used to generate a diversified set of synthetic target images for segmentation training purposes.
- 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 explicitly targets intra-domain variability within the target modality, which is often implicitly ignored in image-level domain adaptation methods.
The avoidance of predefined sub-domain labels such as institute or scanner identifiers makes the approach practical as such labels may often be challenging to acquire reliably in practice.
The visualization and clustering of learned style embeddings provide qualitative evidence that the model captures meaningful appearance variations beyond anatomy.
The method is implemented by extending a rather strong existing image translation backbone rather than redesigning the full pipeline. This allows for a simple incremental contribution.
- 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 novelty of the proposed style modeling is limited relative to prior work in image style transfer and multi-style image translation. The idea of extracting a per-image style code and using it to condition synthesis is well established in the computer vision literature, e.g. https://doi.org/10.1007/978-3-030-58580-8_13 , https://doi.org/10.1109/ICCV48922.2021.01389 The manuscript would benefit from a clearer discussion of what is fundamentally new in the proposed formulation.
The contrastive learning formulation for style extraction appears simplistic and only weakly justified. Positives and negatives are derived from the same target image via spatial cropping and intensity perturbations, which raises questions about what enforces consistency of style representations across different images that share similar acquisition characteristics. Emergent capabilities may arise from such training but this aspect is not well discussed.
The clustering of style embeddings plays a central role in selecting representative styles for downstream segmentation training, yet this step is treated largely as a post hoc procedure. The sensitivity of the results to the clustering method, the number of clusters, or the style selection strategy is not explored, and comparisons against using all target images (thus avoiding clustering entirely) as style references are missing.
- 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
The motivation for emphasizing intra-domain variability would benefit from clearer positioning with respect to modern image translation methods, in particular diffusion-based approaches, which are generally less prone to mode collapse and already capable of generating diverse outputs. A brief discussion of why this issue remains relevant in this context would strengthen the introduction.
Several claims regarding generalizability and robustness would benefit from tighter justification or more precise wording, especially where they rely primarily on qualitative observations, e.g. “the learned style embedding space exhibits meaningful clustering without any supervision” and “the same source image produces visually distinct outputs across references while preserving anatomy”.
The decision to restrict contrastive learning to target-domain images only is not discussed. It would be useful to explain why source-domain images are excluded and whether their inclusion was considered as these would provide additional diverse styles.
Fig. 4 is informative but could be strengthened with additional contextual information, such as indicating institutional or scanner provenance, to support the interpretation of learned styles.
There are occasional grammatical issues and minor phrasing ambiguities that would benefit from careful proofreading to improve readability, especially regarding the use of references as subjects in sentences.
- 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 manuscript addresses a relevant problem of intra-domain variability in medical image domain adaptation for downstream segmentation training and demonstrates measurable gains on a realistic benchmark. The approach is technically sound and integrates well with an established synthesis pipeline.
However, the methodological novelty is limited, with several core ideas closely related to prior work in style-based image translation. Key design choices, particularly the contrastive style learning setup and the clustering strategy, are only weakly justified.
- 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 #4
- Please describe the contribution of the paper
The authors propose IntraStyler, a framework for domain adaptation that does not require site labels to distinguish between imaging sites. The method comprises two components: a contrastive learning module that captures style representations, and a generative framework that synthesizes images in the target domain. Following standard practice, a segmentation network (nnUNet) is then trained on the synthetic images using segmentation labels from the source domain. The authors demonstrate that IntraStyler outperforms existing methods, including the winner of the CrossModa Challenge.
- 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 proposed method is technically sound and novel. By removing the requirement for site codes, which is a limitation of the CrossModa Challenge winner, the approach improves clinical applicability. 2.The use of SLERP is well-motivated, offering finer control over style transfer and reference image generation and is supported by the qualitative figure 3.3.Both quantitative and qualitative results are compelling.
- 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.Evaluation is limited to a single dataset (CrossModa), raising questions about generalizability of the method. 2.The comparison is restricted to a small set of baseline methods, and no ablation studies are provided to assess the contribution of individual components. 3.Several terms in the loss function, specifically L_Adv and L_PatchNCE, are left undefined, which undermines reproducibility and clarity. 4.The caption for Figure 3 should be revised. Describing the output as a smooth transition between two reference images is somewhat misleading, as the interpolated image may adopt the style of the second reference image without fully aligning with its structure.
- 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 core idea of unsupervised domain adaptation without site labels is clinically relevant and well-executed, with results that surpass the CrossModa Challenge winner. However, single-dataset evaluation, no ablation studies, and undefined loss terms limit the strength of the contribution. Despite these limitations, the novelty and practical value makes this an interesting contribution.
- 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
- 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 3D unpaired image translation framework for cross-modality MRI domain adaptation that eliminates the need for predefined site/scanner labels using a contrastively trained style encoder.
Reviewers acknowledge that the paper is addressing a real-world clinical bottleneck: modeling intra-domain variability without relying on coarse sub-domain labels. The contrastive learning strategy is principled, and the method successfully outperforms a highly competitive baseline (the CrossMoDA challenge winner).
To be addressed in final version: 1.Explicitly acknowledge the limitation of evaluating on a single dataset/modality pair (ceT1→T2) and the potential need for modality-specific tuning in other applications.
2.Better position the method relative to prior computer vision literature on style-conditioned synthesis (e.g., StarGANv2, AdaIN) and temper overly broad novelty claims (e.g., “first 3D unpaired…”).
3.Explicitly define all terms in the loss functions (specifically $L_{Adv}$ and $L_{PatchNCE}$).
4.Discuss the impact of the k-means clustering strategy and the uncontrolled number of reference images used (7 vs. 3).
5.Update the captions and context for Figures 3 and 4 as suggested by R3 and R4.
