List of Papers Browse by Subject Areas Author List
Abstract
Deep learning-based 3D medical image segmentation is limited by the high cost of voxel-level annotation. A practical alternative is to label only three orthogonal central slices per volume, yet such sparse 2D supervision provides insufficient constraints to reconstruct topologically consistent 3D structures. Existing methods struggle to bridge this dimensional gap, often failing to segment 2D slices far from the labeled slices, and cannot transfer morphological priors from 2D labeled slices into 3D structures. To overcome these limitations, we propose a novel framework that pioneers the use of the Segment Anything Model 2 (SAM2) as a dimensional bridge. SAM2 has the ability to propagate 2D annotations throughout the 3D volume, but it suffers from propagation drift, i.e., the progressive accumulation of segmentation errors during long-range propagation. To mitigate the drift, we not only propose a cross-view consensus mechanism for adaptively fusing these 2D slice-wise predictions and extending sparse 2D annotations to more robust dense annotations, but also introduce a dynamic 3D refinement module to capture inherent volumetric continuity. Then a confidence-weighted fusion method is designed to obtain the final segmentation results. Experiments on three public benchmarks demonstrate that our method achieves consistent gains (6.7% in Dice on LA2018 dataset) over state-of-the-art baselines. Code is publicly available at https://github.com/DeepMedLab-ECNU/SAM2-bridge.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3034_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/DeepMedLab-ECNU/SAM2-bridge
Link to the Dataset(s)
LA2018 dataset: http://atriaseg2018.cardiacatlas.org/
KiTS19 dataset: https://kits19.grand-challenge.org/
LiTS dataset: https://competitions.codalab.org/competitions/17094
BibTex
@InProceedings{YuPen_SAM2_MICCAI2026,
author = { Yu, Peng AND Chen, Duowen AND Bai, Yunhao AND Wang, Yan},
title = { { SAM2 as a Key Bridge Between 2D Slices and 3D Volumes for Sparsely-supervised Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a novel SAM2-based framework for sparsely supervised 3D medical image segmentation, where only three orthogonal central slices are annotated per volume. The method leverages SAM2 to propagate sparse 2D annotations across the 3D volume, addressing the dimensional gap between 2D supervision and 3D prediction.
To mitigate propagation drift, the authors introduce a cross-view consensus mechanism that adaptively fuses predictions from axial, coronal, and sagittal views. Furthermore, a dynamic pseudo-label refinement module based on a teacher-student architecture is proposed to incorporate volumetric continuity and improve segmentation quality.
The method is evaluated on multiple benchmarks and demonstrates significant improvements over existing semi-supervised approaches.
- 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 a highly relevant and challenging problem in medical image segmentation, namely reducing annotation cost in 3D settings.
- The use of SAM2 as a propagation mechanism is timely and well-motivated.
- The proposed framework is conceptually clear and integrates complementary components (propagation, fusion, refinement).
- The cross-view consensus mechanism is a strong idea that effectively reduces propagation drift.
- The dynamic pseudo-label refinement strategy improves volumetric consistency.
- Experimental results show significant improvements over strong baselines across multiple datasets.
- The evaluation is comprehensive and includes both quantitative and qualitative results.
- The contribution is relevant to the MICCAI community.
- 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 method heavily relies on SAM2, raising the question of how much improvement comes from the proposed framework versus the underlying foundation model.
- The paper does not provide a clear comparison with a direct SAM2-based baseline, making it difficult to isolate the contribution of the proposed modules.
- The overall pipeline is complex, involving multiple stages and components, which may limit practical usability.
- No analysis of computational cost, memory usage, or inference time is provided, which is critical for 3D medical applications.
- The robustness of pseudo-label propagation under challenging conditions (e. g. , noise, anatomical variability) is not thoroughly analyzed.
- Limited discussion on generalization to unseen datasets or domains.
- The method depends on pseudo-label quality, but error accumulation and correction mechanisms are not deeply analyzed.
- 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 paper presents a strong and timely contribution by leveraging foundation models for sparse supervision in 3D medical segmentation. To further strengthen the work, the authors should include a direct comparison with SAM2-only baselines, provide computational analysis, and investigate robustness under challenging conditions.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses an important problem and proposes a well-designed solution that leverages recent advances in foundation models. The performance gains are significant and consistently demonstrated across multiple datasets.
While the method relies heavily on SAM2 and lacks detailed efficiency analysis, the overall contribution is strong and relevant. The paper is above the acceptance threshold and represents a meaningful advancement in sparsely supervised 3D segmentation.
- 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
The primary contribution of this paper is the introduction of the first framework that leverages SAM2 (Segment Anything Model 2) as a dimensional bridge for sparsely-supervised 3D medical image segmentation. The authors developed a unified system comprising three core modules: (1) Utilizing cross-case “global morphological anchors” to execute bidirectional label propagation via SAM2’s inter-frame memory. (2) A slice-level mechanism that adaptively fuses 2D predictions from three orthogonal planes to mitigate propagation drift and single-view bias. (3) A 3D Teacher-Student architecture that progressively restores structural continuity by balancing 2D priors with 3D spatial context.
- 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 method achieves significant performance gains across three public benchmarks (LA2018, KiTS19, and LiTS), notably outperforming the runner-up by 6.7% in Dice score on the LA2018 dataset. (2) The application of SAM2’s powerful memory mechanism to track topological evolution across adjacent medical slices is a novel and effective use of foundation models in a specialized domain. (3) The consensus-weighted fusion avoids hard thresholding, using a soft Dice and intersection area approach to ensure high weights are only assigned to slices where multi-view predictions highly align. (4) By requiring only three orthogonal central slices for annotation, the framework significantly reduces the time-consuming burden of voxel-level labeling in clinical practice. (5) The authors provide comprehensive ablation studies and correlation analyses (e.g., Spearman correlation between consensus and accuracy) to justify the necessity of each component. Additionally, the framework diagram is quite clear.
- 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 effectiveness of the bidirectional propagation relies on “global morphological anchors” constructed by stacking annotated slices across cases. The paper does not extensively discuss how the model handles significant anatomical variations or outliers within these anchor sequences. (2) While SAM2 provides superior propagation, it generally requires more computational resources and memory than traditional lightweight 3D networks like V-Net. A discussion or comparison regarding inference latency and hardware requirements is missing. (3) The weight curve in Figure 1B is too blurry and needs to be improved to publication quality. (4) In the Ablation Study and Qualitative Ablation Analysis sections, if no statistical analysis is conducted, the term ‘significant’ should not be used.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
(1) It claimed to be the first to apply SAM2 to sparse-supervised 3D medical segmentation. The use of global morphological anchors effectively enriches the model’s memory bank across different cases. (2) The Cross-view Consensus-Weighted Fusion elegantly handles SAM2’s propagation drift by using orthogonal views to verify slice-level reliability. (3) The Dynamic Pseudo-label Refinement (DPR) module, utilizing a Teacher-Student V-Net, successfully restores 3D spatial continuity that 2D-slice methods often miss. (4) By proving that “three slices suffice” for accurate 3D volumes, the work provides a practical, low-cost path for training high-performance medical models.
- 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 #3
- Please describe the contribution of the paper
This paper proposes a framework that use the SAM2 as a dimensional bridge between 2D slices and 3D volumes for sparsely-supervised medical image segmentation. Specifically, it not only propose a cross-view consensus mechanism for adaptively fusing these 2D slice-wise predictions and extending sparse 2D annotations to more robust dense annotations, but also introduce a dynamic 3D refinement module to capture inherent volumetric continuity.
- 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 paper presents a well-motivated framework that leverages SAM2 as a bridge between 2D slice-level annotations and 3D volumetric segmentation under an extremely sparse supervision setting (three orthogonal central slices per volume). The key innovation lies in using global morphological anchors constructed from cross-case labeled slices to drive bidirectional propagation via SAM2, effectively generating dense pseudo-labels from minimal 2D supervision. The cross-view consensus-weighted fusion strategy provides a solution to mitigate propagation drift by adaptively weighting slice-wise predictions based on inter-view agreement, while the dynamic pseudo-label refinement within a teacher-student framework further integrates 3D spatial context to restore topological coherence.
- 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 paper claims to be the “first application of SAM2 in sparse-supervised 3D medical segmentation.” However, prior works have already explored SAM-based propagation for sparse annotation scenarios. The authors must clearly differentiate their contribution from these existing efforts or temper the novelty claim accordingly. 2.While SAM2’s memory mechanism is central to the proposed propagation, the paper provides no quantitative comparison against alternative propagation strategies (e.g., simple interpolation, registration-based propagation). Why is SAM2 specifically necessary, and what is the computational overhead incurred by deploying a heavy foundation model for this task? 3.The method uses NL=5 labeled volumes for all datasets, but does not explore how performance scales with varying numbers of labeled cases (e.g., 1, 3, 10). Given that sparse supervision settings are inherently sensitive to the amount of labeled data, this omission significantly weakens the generalizability claims. 4.The cross-view consensus correlation analysis (Fig. 3b) shows strong positive correlations, but does this hold consistently across all slices, including those near volume boundaries where ground truth is absent? What happens when two views simultaneously produce erroneous predictions?
- 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?
My recommendation stems from a careful evaluation across three critical aspects: the soundness of the methodology, writing quality, and experimental thoroughness.
- 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.
The reviewers acknowledged the methodological contributions in 3D medical segmentation with sparse supervision and thorough empirical evaluation. The authors are encouraged to address the concerns regarding computational cost, missing baseline comparison, lack of discussion of generalization, etc. The authors are also suggested to make their codes public.