List of Papers Browse by Subject Areas Author List
Abstract
Foundation models provide strong anatomical priors that can improve semi-supervised 3D MRI segmentation under limited annotation. However, many approaches still rely on unstable teacher-student updates and heuristic pseudo-labeling, which can accumulate errors when learning from unlabeled volumes. In MRI, scanner-dependent intensity inhomogeneity and distribution shifts further degrade robustness, especially in extreme low-label regimes with heterogeneous lesions. We propose ReLiF-3D, a foundation-guided semi-supervised framework that trains a lightweight 3D U-Net under supervision from a frozen volumetric foundation model (SAM-Med3D) prior, avoiding EMA teacher drift while retaining a stable spatial guide. Our main contribution is the Smooth Orthogonal Bias Field (SOBF), a paired-view generator that simulates realistic MRI bias through orthogonal multiplicative and additive smooth fields with a progressive strength schedule, producing anatomically consistent yet scanner-shifted views for reliable consistency learning. To stabilize unlabeled optimization, ReLiF-3D applies confidence-gated voxel co-regularization to enforce consistency only on reliable agreements and lesion-aware representation alignment guided by the SAM-Med3D prior. Experiments on public BraTS and Left Atrium datasets with 1-5 labeled volumes demonstrate significant improvements in overlap and boundary metrics over SAM-assisted SSL baselines and state-of-the-art methods. Project page is available at https://visdomlab.github.io/ReLiF_3D
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1855_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/kunaljangid01/ReLiF_3D
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JanKun_ReLiF3D_MICCAI2026,
author = { Jangid, Kunal AND Basu, Tanmay AND Kurmi, Vinod},
title = { { ReLiF-3D: Prior-Guided Semi-supervised 3D MRI Segmentation via Robust Bias-Consistent Paired Views } },
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
1.This paper proposes a semi-supervised framework for 3D MRI segmentation under limited annotations, which leverages a frozen foundation model (SAM-Med3D) as a stable supervisory prior to replace the conventional EMA teacher and reduce training instability. 2.It further introduces a bias-field-inspired augmentation (SOBF) and reliability-aware consistency mechanisms (CGCR and LARC) to improve robustness and enhance lesion representation learning in low-label settings.
- 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 addresses an important problem of semi-supervised 3D MRI segmentation. A strength is the use of a frozen foundation model as a stable supervision signal instead of EMA teacher. The SOBF augmentation is well motivated for MRI, and the proposed consistency designs help improve robustness in low-label settings. The method shows consistent improvements over strong baselines.
- 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 overall novelty is moderate, since the method mainly combines existing ideas such as foundation-model guidance, bias-aware augmentation, and confidence-based consistency learning. 2.The ablation study is not fully sufficient to clearly isolate the contribution of each component, especially the frozen SAM prior versus the auxiliary modules. 3.The method may rely heavily on the quality of SAM-Med3D, but the paper does not analyze this dependency or its robustness under domain shift. 4.The experimental validation is still limited, as cross-dataset or cross-scanner generalization is not explicitly evaluated.
- 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?
The paper addresses a relevant problem and shows consistent improvements, but the novelty is moderate as it mainly combines existing ideas. In addition, the ablation and analysis are not fully sufficient to clearly support each component. Overall, the contribution is useful but not strong enough for a higher score.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes ReLiF-3D, a foundation-guided semi-supervised framework for 3D MRI segmentation under extreme label scarcity. It couples a frozen SAM-Med3D as a stable prior with a learnable 3D U-Net, avoiding EMA teacher drift. The key contribution is the Smooth Orthogonal Bias Field (SOBF), a paired-view generator that simulates MRI-specific scanner bias through orthogonal multiplicative and additive smooth fields. Experiments are conducted on BraTS and Left Atrium datasets.
- 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.
- SOBF is grounded in MRI bias-field physics rather than generic augmentation, which is a principled approach to addressing scanner variability. The orthogonality constraint between multiplicative and additive fields is a nice design to ensure non-redundant perturbations. 2.Using a frozen SAM-Med3D as the supervisory prior instead of an EMA teacher is a sound design choice, particularly well-justified in the extreme low-label regime where EMA targets can be unreliable. 3.The paper compares against six SSL baselines combined with SAM-Med3D across four label settings on two datasets, with statistical significance testing (Wilcoxon test), which strengthens the experimental rigor.
- 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 authors state the challenge of scanner-dependent intensity inhomogeneity and distribution shifts of MRI. However, both BraTS and LA are well-studied benchmarks with relatively clean data are single-sequence single-scanner MRI. Testing on more challenging real-world datasets with actual scanner variability like M&Ms [1] would further explore the method. 2.To improve clarity, I recommend splitting the combined table 2 focused on separate experimental themes, or add clearer, standalone caption. This avoids mixing SOTA comparisons, module ablations, and loss/prompt ablations within a single layout. 3.Running a frozen SAM-Med3D per iteration for every input volume is computationally expensive. The paper does not report training time comparisons with baseline methods. 4.The progressive scheduling of supervised loss (interpolating Dice and CE) is presented as a contribution but is a common practice in medical segmentation. The authors should tone down this claim. 5.The author did not indicate that they would open source code. The reproducibility is questionable
[1] Campello, Victor M., et al. “Multi-centre, multi-vendor and multi-disease cardiac segmentation: the M&Ms challenge.” IEEE Transactions on Medical Imaging 40.12 (2021): 3543-3554.
- 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?
A solid contribution with a well-motivated MRI-specific augmentation strategy. The frozen foundation prior is a reasonable design choice. However, the experiments are limited to well-studied benchmarks and lack of evaluation on diverse data with actual scanner variability. The author did not indicate that they would open source code. The applicability in more challenging real-world MRI datasets is questionable.
- 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 paired-view generator (SOBF) is a novel design. The authors creatively propose combining an orthogonal multiplicative field with an additive field to simulate realistic MRI bias effects, coupled with a progressive strength schedule to produce paired views that preserve anatomical consistency while exhibiting scanner-distribution-shift characteristics. Compared with generic intensity jittering, this is more task-specific and provides a solid foundation for reliable consistency learning. The overall framework design is well-motivated. The introduction of Confidence-Gated Co-Regularization (CGCR) and Lesion-Aware Representation Consistency (LARC) effectively prevents error propagation during unlabeled-data optimization and, under the guidance of the SAM-Med3D prior, focuses feature learning on lesion regions. The three modules work synergistically to achieve significant performance gains in extreme low-label scenarios.
- 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 experimental results are convincing. Consistent advantages are demonstrated on both the BraTS and LA datasets under the extreme setting of 1–5 labeled samples. With only a single labeled sample, Dice scores reach 76.41% and 76.84% respectively, significantly outperforming all compared baselines (as indicated by the † markers in Table 1 denoting Wilcoxon significance test results).
- 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 authors claim that the work aims to address the cross-domain generalization problem caused by bias fields. However, the experiments only apply synthetic intensity perturbations on the same dataset to simulate domain shift, with no true cross-site validation. This is insufficient to demonstrate the model’s generalization ability when facing genuinely different scanners and acquisition protocols in practice. Compared with classical approaches such as polynomial basis function bias-field simulation, which is grounded in physical principles (e.g., TorchIO’s RandomBiasField), the authors’ Gaussian noise plus box-filtering scheme lacks the necessary theoretical justification. The paper does not explain the rationale or advantages of this scheme for modeling the true distribution of MRI bias fields. The introduction of the additive field together with the orthogonality constraint constitutes the main novelty of SOBF relative to traditional methods. However, the authors neither compare against a standard polynomial multiplicative field as a baseline, nor include a non-orthogonal dual-field combination as a key control. It is therefore difficult to conclusively demonstrate the practical gain of the box-filtered, orthogonally-designed dual-field scheme over existing methods. Additional ablation experiments are recommended to support this design choice.
- 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 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.
(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 paper’s core contribution, SOBF, delivers consistent and significant performance gains under extremely limited annotation, and the overall framework is well-designed with sufficient experimental validation, which clearly meets the acceptance threshold. However, although the authors motivate their method as addressing cross-scanner generalization, no true cross-site validation is provided, and key control experiments isolating the benefit of SOBF over existing bias-field modeling approaches are missing, making it difficult to fully attribute the observed gains to the proposed design. Overall, the contributions are valuable but the justification remains incomplete, hence my Weak Accept recommendation.
- 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.
All reviewers rated the paper above the threshold (consistent weak accepts) and agree it tackles an important problem with a well-motivated and practically useful framework. Key strengths include the use of a frozen foundation model as a stable supervisory prior, the MRI-specific SOBF augmentation grounded in bias-field modeling, and consistent, statistically validated improvements, especially in extremely low-label settings where gains are substantial. While reviewers note moderate novelty and some missing analyses (e.g., deeper ablations, cross-domain validation), these are seen as limitations rather than flaws and do not undermine the empirical effectiveness or practical relevance of the approach. Overall, the paper provides a solid and useful contribution that meets the acceptance bar.