Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Self-supervised learning enables pretraining on large unlabeled 3D brain MRI cohorts. However, many objectives are adapted from natural images and underuse neuroanatomical spatial structure. We build on an iBOT-style teacher–student 3D vision transformer (ViT) with an exponential moving average (EMA) teacher. We align masked token representations in latent space. We also learn subject-level discrimination from classification tokens (CLS). To inject spatial priors, we propose distance-modulated patch decorrelation. It suppresses spurious similarity more for far-apart patches. It remains tolerant to nearby regions. We pretrain on 6,811 unlabeled T1-weighted scans from five independent cohorts. These cohorts do not overlap with the downstream datasets. We evaluate transfer on AD-related classification using the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies 3 (OASIS-3), with AD vs. NC evaluated on both datasets and MCI vs. NC evaluated on ADNI only. We also test tumor grading on Brain Tumor Segmentation (BraTS). Beyond downstream metrics, we analyze distance-dependent patch similarity. We also explain why region of interest (ROI)-based priors can be unreliable at our patch granularity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3889_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/MinKyeong-98/Brain-dis-ssl/tree/main

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimKye_BrainSpecialized_MICCAI2026,
        author = { Kim, Kyeong Ho AND Kwon, Minkyeong AND Kim, Yu-Mi AND Kim, Mi Kyung AND Shin, Min-Ho AND Chung, Insung AND Koh, Sang Baek AND Kim, Hyeon Chang AND Lee, Jong-Min},
        title = { { Brain-Specialized Self-supervised Learning: Neuroanatomy-Aware Spatial Priors for 3D Brain MRI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a brain-specialized self-supervised learning framework that incorporates neuroanatomical spatial structure as an inductive bias. The primary methodological contribution is the distance-modulated patch decorrelation prior, which suppresses spurious similarities between spatially distant patches while remaining tolerant of nearby regions. This is integrated into an iBOT-style teacher-student Vision Transformer that combines latent-space masked token alignment and global subject-level discrimination.

  • 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 study appropriately identifies the limitations of standard self-supervised learning objectives for natural images when applied to brain MRI, and proposes a spatial prior that better captures the anatomical regularity and low-texture characteristics of this domain.
    • The model is pretrained on a large-scale unlabeled dataset of 6,811 T1-weighted scans collected from multiple cohorts, which supports strong generalizability.
    • The training framework is computationally efficient, as it aligns representations in the latent space without relying on a heavy decoder for pixel-level reconstruction.
    • The approach is thoroughly evaluated across multiple benchmarks, and the analysis provides useful insights into why ROI-based priors may fail at the selected patch level due to mixed-region tokens and label noise.
  • 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.The distance-based weighting is implemented as a simple linear function, which may not adequately capture complex neuroanatomical relationships or long-range dependencies. A more flexible non-linear formulation could potentially improve the modeling of spatial interactions. 2.The framework is restricted to T1-weighted MRI, and its applicability to other important modalities such as FLAIR, T2, or PET is not examined, limiting its broader relevance. 3.The treatment of symmetric anatomical regions remains unclear. Although distant patches are decorrelated, the method does not explicitly address whether this negatively impacts regions that are anatomically similar but spatially separated, such as bilateral structures. 4.The analysis of patch size is limited, as the study does not include a systematic evaluation of how different patch resolutions influence the effectiveness of the proposed distance-based prior.

  • 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.

    (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 work successfully adapts the latest advancements in self-supervised learning to the specific constraints of 3D neuroimaging. The distance-modulated decorrelation is a clever, annotation-free spatial cue that significantly improves patch discriminability. Supported by large-scale pretraining and a diverse set of downstream tasks, this paper provides a robust contribution to the field of medical image representation learning.

  • 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

    To address the limitations of existing self-supervised learning (SSL) methods for 3D brain MRI, particularly their insufficient utilization of neuroanatomical spatial structure and the tendency for brain-region patch features to become ambiguous, this work proposes a hybrid SSL framework tailored to brain imaging.

    Built upon an iBOT-style teacher-student 3D Vision Transformer (ViT) architecture, the method integrates latent-space masked token alignment with CLS token contrastive learning. Moreover, it introduces a novel spatial prior termed Distance-Modulated Patch Decorrelation (DIS), which replaces conventional pixel reconstruction objectives and ROI-based priors. DIS explicitly encourages greater discriminability between spatially distant patches, leveraging intrinsic anatomical structure.

    The model is pretrained on 6,811 unlabeled T1-weighted MRI scans. Downstream evaluations are conducted on Alzheimer’s disease classification (ADNI, OASIS-3) and brain tumor grading (BraTS). Results show that DIS effectively enhances long-range patch discrimination and yields significant improvements in tasks with subtle disease signals, such as MCI vs. NC classification and tumor grading. In contrast, ROI-based priors prove ineffective due to the mismatch between coarse anatomical regions and fine-grained patch representations.

  • 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 Distance-Modulated Patch Decorrelation (DIS) introduces no additional annotation requirements and incurs minimal computational overhead, making it well suited to the abundance of unlabeled brain MRI data. It offers a more flexible and scalable alternative to traditional ROI-based anatomical priors.

    2.The effectiveness of DIS is clearly demonstrated through analyses of patch similarity decay over spatial distance and ROI purity evaluation, which also explain why ROI-based priors fail under fine-grained patch representations.

    3.Notably, the method shows pronounced improvements in subtle-signal tasks such as MCI classification and brain tumor grading, aligning well with the needs of early-stage clinical diagnosis.

  • 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.Marginal gains on the core task. On the primary AD vs. NC classification task, the method shows no improvement over iBOT. The advantages are mainly observed in MCI vs. NC and BraTS tumor grading, raising questions about its general effectiveness on more established benchmarks.

    2.Lack of statistical significance testing. No statistical significance analysis (e.g., p-values) is reported when comparing the proposed method with baselines. Without such tests, it is unclear whether the observed improvements are statistically meaningful or simply due to random variation.

    3.Incomplete ablation studies. The paper does not conduct combinatorial ablations of different loss components, making it difficult to disentangle the independent contributions and potential synergistic effects of each module.

    4.No hyperparameter sensitivity analysis. The loss weights are both fixed at 1, without investigating their impact on performance. In addition, the masking ratio and EMA momentum are set to default values, and no parameter sensitivity experiments are provided to validate the robustness of these choices.

  • 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 provide sufficient information for 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.

    (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?

    See the weaknesses.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Thank you for the detailed rebuttal and for clarifying several points.

    However, some key concerns remain. While I appreciate the more cautious interpretation of the results, the lack of improvement on the primary AD vs. NC task still limits the claim of general effectiveness. The reported gains on other benchmarks are encouraging, but without formal statistical significance testing, it is unclear whether these improvements are robust.

    Regarding ablations, I understand the rationale for the progressive design, but the absence of more complete combinatorial analysis makes it difficult to isolate the specific contribution of DIS. In addition, no hyperparameter sensitivity analysis is provided, leaving uncertainty about the robustness of the reported performance to loss weights and patch size choices.

    Finally, although the controlled baseline setting is reasonable, the limited comparison scope and lack of segmentation or other localization-sensitive evaluations constrain the strength of the representation-level claims.

    Overall, the rebuttal clarifies the positioning of the method, but concerns about statistical validation, ablation completeness, and robustness remain. Strengthening these aspects would improve the empirical rigor of the work.



Review #3

  • Please describe the contribution of the paper

    This manuscript proposes a brain specific SSL framework for 3D T1 MRI built around an iBOT-style teacher–student ViT. They combined latent masked-token alignment, CLS-level discrimination, and a new distance modulated patch decorrelation objective that weights repulsion more strongly for spatially distant patches. Authors’ main claim is that explicit 3D spatial priors can improve patch distinctiveness in brain MRI and yield better transfer to downstream tasks such as AD diagnosis and tumor grading.

  • 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.

    This manuscript uses a distance-aware regularizer that uses 3D patch-center geometry to modulate patch decorrelation, which is a natural fit for volumetric neuroimaging where spatial organization matters.

    The distance weighted decorrelation term is lightweight and easy to add to existing teacher–student ViT pipelines. This simplicity is interesting compared with heavier pixel reconstruction or atlas-dependent supervision. Specially, the equation in Section 3.4 is straightforward and likely reproducible in principle.

    Authors evaluated their work on ADNI AD/NC, ADNI MCI/NC, OASIS-3 AD/NC, and BraTS grading. The results are not good enough, but this comparison is interesting.

    The ablation (Table 2) is the strong part of this manuscript. It isolates the effects of masked token learning, CLS discrimination, the proposed distance prior, and an ROI-based alternative. This Table 2 supports that the spatial term contributes beyond the base hybrid SSL setup.

    Rather than showing downstream results, the authors analyze distance-dependent patch similarity and investigating why atlas-based ROI priors fail at the chosen 141414 patch granularity. Specially, Fig 3 shows a plausible explanation for why a seemingly more anatomical prior can degrade performance. This analysis is a good addition.

  • 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.

    SSL is rapidly evolving, but the related work section does not sufficiently discuss recent papers. It is unclear why more up-to-date methods, especially in 3D medical SSL, are not included. The authors should better justify their choice of references and clearly position their work with respect to the latest developments in this area.

    Authors did not share their code for reproducibility. In a field like SSL, where implementation details can significantly affect performance, it is very important to provide code along with proper comments and guidelines. This would allow others to reproduce the results and build upon the work.

    The paper has compared the results with set of baselines. While the paper includes standard methods such as MAE, SimMIM, and iBOT, it does not fully evaluate against several relevant recent SSL approaches specifically designed for 3D medical imaging. Recent literature continues to show rapid progress in this area.

    All downstream tasks are classification tasks. But this manuscript centered on patch-level spatial structure in 3D MRI. Need a clear argument why they only chose only classification as downstream tasks and not included tasks, which are local representation quality matters more directly, such as segmentation, fine-grained region-level analysis.

    The results in Table 1 also raise some concerns. The improvements over strong baselines are fairly modest and inconsistent across settings, which makes it difficult to assess the true value of the proposed approach. The authors should do more to explain these gaps and make a clearer case for why their method matters. It is still unclear why this work is necessary and whether it represents a meaningful step forward or a new direction for SSL.

  • 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.

    (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 represents a sensible, domain-aware contribution. In addition to that, the empirical evidence supports that the distance-modulated spatial prior is beneficial, mainly on harder transfer setups like MCI and BraTS. Nevertheless, the gains are insufficient to be absolute due to the lack of capacity. However, this method benefits over strong modern SSL baselines but is not enough for a higher score.

    The strongest reason to accept is that the paper identifies a real mismatch between generic SSL objectives and brain MRI structure, after that they have proposed a lightweight, anatomically informed fix that is supported by ablations and representation analysis. The main reason not to score it higher is that the comparison set is not comprehensive enough for 3D medical SSL.

  • 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

We sincerely thank the reviewers for their constructive comments. 1.Empirical evidence and statistical support(Rmeta, R2, R3) We acknowledge that DIS does not uniformly outperform iBOT across all benchmarks and does not improve over iBOT on ADNI AD/NC. Our results should therefore be interpreted as showing task dependent benefits rather than uniform superiority. DIS is a patch level spatial regularizer designed to reduce spurious similarity among distant patches. Therefore, its benefit is expected to be clearer when disease cues are subtle, spatially distributed, or heterogeneous. In line with this interpretation, DIS improves over iBOT on ADNI MCI/NC from 0.69 to 0.71 AUC and on BraTS from 0.59 to 0.73 AUC. Although performance was evaluated using 5-fold cross-validation, formal statistical significance testing between methods was not performed. Therefore, we do not claim statistically significant superiority and interpret the observed gains conservatively. 2.Ablation design and hyperparameter sensitivity(Rmeta, R1, R2) Regarding ablation, the masked token objective is the base pretext task, not an optional auxiliary loss. It creates the masked student/unmasked teacher asymmetry. Without it, the two views would be nearly identical under the same augmentation pipeline, making the CLS objective less informative. Since DIS is defined on patch representations learned in this masked view setting, removing masked token learning would change the pretraining objective rather than isolate DIS. Thus, Table 2 uses a progressive ablation design. This design tests whether DIS provides additional benefit beyond the base SSL objectives and whether a distance-aware prior is preferable to an atlas derived ROI prior at the selected patch granularity. For controlled comparison, the masking ratio and EMA momentum followed standard MAE/iBOT-style settings and 14×14×14 patches were used to balance resolution, token length and computation. Sensitivity to λ_cls, λ_sp and patch size was not exhaustively explored, which is a limitation of the submitted study. 3.Baseline scope and downstream evaluation(Rmeta, R3) We acknowledge the limited baseline scope and absence of segmentation evaluation. Our experiments were designed to isolate the proposed spatial prior under matched backbone and training conditions, rather than to provide an exhaustive benchmark of all recent 3D SSL systems. MAE and SimMIM represent masked image modeling baselines, while iBOT is the closest teacher-student token level SSL framework to DIS. This comparison allows the effect of the distance-modulated spatial prior to be examined under a controlled setting. We also agree that segmentation would more directly assess local representation quality. In the submitted study, we focused on classification to evaluate the pretrained encoder under controlled transfer settings, because segmentation performance can be affected by decoder design, resolution recovery and task specific optimization. The submitted patch-level similarity decay and ROI purity analyses provide representation-level evidence that DIS changes patch representations as intended. 4.Spatial prior design and modality scope(R1) We used a simple linear distance weight as a minimal, monotonic and parameter free spatial prior. The goal is not to model nonlinear anatomy or bilateral symmetry, but to reduce spurious similarity between distant patch tokens in low texture 3D MRI. DIS uses soft distance-modulated decorrelation rather than hard negatives, so homologous regions can remain similar when supported by image evidence and the base SSL objectives. The ROI prior degraded performance, suggesting that rigid atlas constraints can be noisy at 14×14×14 patch granularity due to mixed region patches. Evaluation on FLAIR, T2, PET, explicit symmetry modeling and learnable distance priors were outside the scope of the submitted experiments. 5.Reproducibility To support reproducibility, we plan to release the source code upon acceptance.



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 overall recommendation from the three reviewers is borderline, with one weak accept and two weak rejects. The authors need to make rebuttal to address the following concerns: First, the empirical evidence presented in the paper is unconvincing for the majority of reviewers: performance gains over strong baselines are modest and inconsistent, with no improvement on the primary AD vs. NC classification task, and no statistical significance testing is reported. Other major concerns include an incomplete comparison with recent 3D medical SSL methods, the absence of evaluation on segmentation tasks that would directly reflect patch-level representation quality, and insufficient ablation of individual loss components and hyperparameter sensitivities.

  • 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 paper proposes a neuroanatomy-aware self-supervised learning framework with a distance-modulated patch decorrelation prior. While the initial reviews were mixed, the author rebuttal effectively clarified the scope of the contributions and the limitations. After rebuttal, Reviewer #2 raised their score to Accept, joining Reviewer #1’s Weak Accept. After reading the paper, I think that the strengths of this work lie in a lightweight, well-motivated spatial regularizer that demonstrably improves patch discriminability in low-texture 3D brain MRI, particularly on subtle-disease tasks like MCI classification and tumor grading. The ablation and representation analyses also convincingly explain why the proposed prior outperforms rigid atlas-based ROIs. The rebuttal’s conservative interpretation of results and commitment to code release further strengthen the case. Overall, the paper provides a domain-appropriate inductive bias that advances representation learning for brain MRI, which I think can be considered as acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The paper correctly identifies a key limitation of generic SSL for 3D brain MRI and proposes a lightweight, effective spatial prior (DIS) that improves performance on subtle-signal tasks such as MCI classification and tumor grading. All three reviewers gave scores of 4, 3, and 3 (no Strong Reject), and the authors’ rebuttal honestly acknowledges the method’s task-dependent benefits. The remaining concerns are addressable and do not undermine the core contribution.



Meta-review #3

  • 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.

    After considering the rebuttal and updated reviews, I lean toward rejection. The rebuttal clarifies the intended scope of the contribution; however, several important concerns remain insufficiently addressed, including the lack of statistical validation, sensitivity analysis, limited downstream evaluation, and only task-dependent rather than consistent improvements over strong SSL baselines. Therefore, I do not think the paper sufficiently crosses the acceptance threshold in its current form.



back to top