List of Papers Browse by Subject Areas Author List
Abstract
3D Automated Breast Ultrasound (ABUS) is an effective imaging modality for breast cancer detection. However, tumor segmentation within this modality remains challenging due to the diversity of tumor characteristics. While foundation models (e.g. SAM and SAM2) offer a generalizable approach, their performance degrades in medical imaging due to domain gaps. Several SAM adaptation methods have been proposed to mitigate this issue, but they rely on costly pixel-level annotations and struggle to adapt to unseen medical domains such as 3D ABUS with limited labeled data. To this end, we propose a label-efficient adaptation of SAM2 using point annotations, significantly reducing the need for full pixel-level labels. Given a point label, SAM2 can perform 3D segmentation from 2D images sliced from the 3D volume along three orthogonal axes. Since the slices from different axes exhibit distinct characteristics, which are complementary to each other, we introduce multi-axis consistency regularization. By employing pseudo labels predicted from the other axes, the model is trained to learn complementary information, thereby improving accuracy even with less labeled data. We evaluate our method on two different datasets and ours trained with point labels outperforms existing methods trained with a large amount of full labels.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1627_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{KimSoo_Multiview_MICCAI2026,
author = { Kim, Soopil AND Lee, Dongmin AND Jang, Jiwon AND Arsalane, Wafa AND Kim, Daekyung AND Lee, Junhyeok AND Lee, KyoungJoon AND Park, Sang Hyun},
title = { { Multi-view Consistency-Based Adaptation of SAM2 for 3D ABUS Tumor Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
I think the main contribution of this work is the adaptation of SAM2 (a video-based foundation model) to 3D medical segmentation via slice-wise processing and the use of point annotations rather than dense masks.
- 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 introduces a multi-axis consistency regularization tailored to SAM2’s sequential propagation to adapt 3D volumes from single-point prompts. – There is a learnable detection token to estimate tumor presence per slice, and integrating these probabilities into cross-axis pseudo-label fusion is a simple yet effective idea to suppress noisy pseudo-labels. –Method pipeline figures help clarify the architecture components (frozen encoder, learnable adapters, axis-specific decoders).
- 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 major concern is that cropping to tumor-containing regions for all methods may require additional supervision at test time. The chosen procedure and its fairness across baselines are not sufficiently explained. How are the tumor-containing ROIs cropped at test time? Is any ground-truth information used to localize the tumor? –I believe there is limited Novelty at the conceptual level. Core components are incremental combinations of SAM2 adaptation, Consistency regularization, and Pseudo-labeling. –Also, there is no formal analysis of why multi-axis consistency works. I have few questions for the authors: –Are the point annotations in TDSC-ABUS real or simulated from ground-truth masks? If simulated, can you report results with truly independent point annotations to validate label-efficiency claims? –How do the learned detection token predictions calibrate (e. g. , PR/ROC per slice)? Did you attempt uncertainty weighting or confidence-thresholding for pseudo-label fusion?
- 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?
While the paper presents a practical and well-engineered approach for weakly-supervised 3D medical segmentation using SAM2, the methodological novelty is limited, as it largely combines existing ideas such as pseudo-labeling and multi-view consistency. The lack of theoretical insights, limited experimental validation, and insufficient comparison with strong baselines weaken the overall contribution. Although the empirical results are promising, the work does not meet the bar of conceptual innovation and depth typically expected at a top-tier conference.
- 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
The proposed method shifts away from expensive full-mask supervision, opting for a weakly-supervised learning (WSL) approach that utilizes limited labeled data and a larger set of simple point annotations.
- 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.
Innovation in WSL, where this may be the first study applying weakly-supervised learning based on SAM for 3D medical segmentation. Practicality, where by reducing the reliance on pixel-level labels (which require expert radiologists) in favor of point prompts, the method is highly scalable for clinical dataset construction. Architectural Fit, where leveraging SAM2’s temporal/sequential processing for 3D volumetric slices is a clever and effective use of the foundation model’s inherent strengths.
- 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.
Computational Overhead, where training three separate axis-specific models and performing multi-round pseudo-labeling likely requires significant GPU resources ($3 \times$ NVIDIA A6000 were used). Resolution Dependency, where the authors noted the z-axis performed best due to its higher resolution, suggesting performance might degrade in imaging modalities with highly anisotropic voxels.
- 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.
(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?
This paper is a very strong candidate for acceptance due to its clear clinical utility and technical innovation. It has shown the following highlights: significant performance gains, label efficiency, technical novelty, rigorous evaluation etc.
- 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 authors introduce the method of automated breast ultrasound (ABUS) imaging relying on the SAM2.
The authors use TDSC-ABUS and private dataset for the training and evaluation.
The authors claim achieving significant performance gains with as low as 20% of labels.
The authors compare the method to some others in the state of the art, showing the advantages of the method.
Overall, the presented metrics appear appealing, making a method of interest not only for ABUS, but for other 3D modalities.
- 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.
3D ABUS is a recent, clinically relevant imaging modality making the paper of interest to the field.
Using SAM in weakly supervised setting appears to be a nifty hint allowing to achieve high performances with few to none annotations.
The authors were able to fit in several ablation studies, exploring the proposed method from several angles.
- 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 appear to lack comparison with state of the art. Better context would have helped to understand the contributions. That is, it is not perfectly clear how the paper is novel in the area of ABUS-related methods and how does it compare.
- 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
State of the art: It would be helpful if the authors put some light on the state-of-the-art performances. In particular, [1] state the performances on the TDSC-ABUS dataset.
Methods The paper may benefit from improving clarity: in particular,
- the weakly supervised has substantial amount of notations, I wonder if the authors may make it clearer
- the idea of multi-round training seems a bit confusing. I wonder is the authors can revise the presentation? In particular, what is the order of rounds?
Experiments I wonder if the authors can comment or detail more how the resolution of images varies and how may it affect the pre-processing?
Results The authors present the multi-round training. Could the authors comment of provide more details about running further rounds?
Overall Finally, I suggest the authors to proof-read the manuscript to avoid typos and
[1] https://www.nature.com/articles/s41598-025-97098-w.pdf
- 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 authors address a clinically and technically relevant task. However, the manusxript does not offer sufficient context for a good understanding of the advantages of the proposed method.
- 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
We thank the reviewers and the meta-reviewer for the thoughtful and constructive comments.
(1) Regarding ROI cropping, our task is point-prompted 3D tumor segmentation, not fully automatic tumor localization. At inference, the only test-time prompt is a single foreground point. The 256×256×128 ROI is a fixed-size crop centered on this point and resized to 128×128×128 for computational feasibility. No ground-truth mask, boundary, or tumor-size information is used. The same crop is applied to all compared methods; thus, it does not introduce dense-label leakage or favor our framework.
(2) Regarding point annotations, the WSL setting assumes center-point labels for weakly labeled volumes. For TDSC-ABUS, where dense masks are available, point prompts are simulated from tumor centers of training masks. Dense masks are used only for the fully labeled subset and final evaluation. This protocol directly tests whether most dense 3D masks can be replaced by point-level prompts, and it is consistently applied to all WSL baselines.
(3) Regarding novelty, while pseudo-labeling, consistency regularization, and SAM adaptation have precedents, our contribution is not a generic combination of them. We formulate the first weakly supervised adaptation of SAM2 for 3D medical segmentation by exploiting SAM2’s sequential propagation across orthogonal 2D slice sequences. Three axis-specific SAM2 adapters generate complementary 3D predictions from the same point prompt, while a slice-level detection token estimates tumor presence and weights cross-axis pseudo-label fusion. This design is specific to SAM2-based 3D volumetric adaptation and differs from dense-mask SAM adaptation, SAM-Med3D-style supervised 3D adaptation, and conventional SSL methods that do not exploit multi-axis SAM2 propagation.
(4) The ablation study supports the design. Naïvely reusing same-axis pseudo-labels degrades performance, whereas using predictions from other axes improves DSC. Detection-score-weighted fusion further improves performance, and multi-round cross-axis training provides additional gains before saturation. These results show that the improvement comes not merely from extra weak labels, but from complementary multi-axis guidance and detection-aware pseudo-label construction.
(5) Regarding SOTA comparison, the paper evaluates supervised, semi-supervised, weakly supervised, SAM-based, and SAM2-based baselines under the same split, crop size, prompt setting, and label-budget protocol. Baselines include 3D U-Net with SSL methods, SAM-Med3D, CPC-SAM, MedSAM, SAM2 zero-shot, and SAM2-Adapter variants. Across TDSC-ABUS and the in-house dataset, our method consistently achieves the best DSC; notably, using only 20 fully labeled volumes, it outperforms several baselines trained with all 80 fully labeled TDSC-ABUS volumes, supporting the label-efficiency claim.
(6) Regarding reproducibility and multi-round training, the paper specifies the SAM2 hiera-large backbone, trainable adapters and mask decoder, AdamW optimizer, learning rate, cosine decay, epochs, pseudo-label threshold, focal-loss parameters, and three WSL rounds. The training order is: axis-specific models are first trained with the fully labeled subset; in each WSL round, predictions and detection scores from the two non-target axes are fused to supervise the target-axis model; updated models are then used in the next round. The final model uses the best-performing single axis rather than an ensemble, so the multi-axis design improves WSL training without increasing inference-time complexity.
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 presents a weakly-supervised framework for 3D medical image segmentation by adapting the SAM2 to volumetric data using point-based annotations and multi-axis consistency regularization. The reviewers highlight the practical clinical utility of the approach, the effectiveness of leveraging SAM2’s sequential processing for 3D volumes, and the promising performance gains achieved with limited labels. However, concerns were raised regarding the conceptual novelty of the method, the potential for information leakage during the ROI cropping process, and the lack of comprehensive comparisons with existing state-of-the-art methods. Furthermore, the clarity of the multi-round training procedure and the reproducibility of the results require further justification. Given the divergent scores (6, 4, 3), the authors are encouraged to provide a detailed rebuttal addressing the fairness of the experimental setup, the specifics of the point-annotation simulation, and the methodological distinctions from prior work.