List of Papers Browse by Subject Areas Author List
Abstract
Deep learning-based medical image segmentation typically relies on massive, high-quality expert annotations, which are costly and time-consuming to obtain. Cold-start Active Learning (CSAL) alleviates this burden by querying informative samples for annotation when dealing with a new unlabeled training set. However, current CSAL methods either suffer from selection redundancy and boundary outliers bias, or rely on coarse heuristics that poorly approximate true segmentation difficulty. Meanwhile, foundation models like Segment Anything Model 3 (SAM3) with zero-shot segmentation ability can provide useful structural cues for CSAL, yet directly applying them to medical images leads to unstable mask quality. To this end, we propose a unified Dual-Space CSAL framework that leverages SAM3 outputs as zero-cost priors for sample selection, while simultaneously capturing highly informative samples that offer both feature-level typicality and mask-level diversity. In the image feature space, we introduce Centroid Affinity Scoring coupled with a Cluster-Balanced Penalty to explicitly quantify sample typicality, preventing redundant queries within high-density cores. Furthermore, to eliminate mask-level redundancy across diverse image features, we introduce a pseudo-mask descriptor space with a Morphological Diversity mechanism, which utilizes uncertainty-weighted shape priors to ensure maximum shape coverage within the query set. Extensive experiments on PROMISE12 and ISIC 2018 datasets demonstrate that the proposed framework significantly outperforms state-of-the-art CSAL methods, reaching performance comparable to the fully supervised upper bound with only about 5% labeled samples on the skin lesion segmentation task. Code is available at https://github.com/HiLab-git/DSAS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4655_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HiLab-git/DSAS
Link to the Dataset(s)
ISIC 2018 dataset: https://challenge.isic-archive.com/data/
PROMISE12 dataset: https://zenodo.org/records/8026660
BibTex
@InProceedings{YePin_DualSpace_MICCAI2026,
author = { Ye, Ping AND Zhu, Ning AND Fu, Jia AND Zhou, Xianhao AND Luo, Zihao AND Wang, Litingyu AND Li, He AND Wang, Guotai},
title = { { Dual-Space Cold-Start Active Learning Guided by SAM3 for Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents an innovative combination of representative-based and coverage-based methods for sample selection. By leveraging masks generated by the zero-shot foundation model SAM3 (via text prompts) to construct a diversified description space, the proposed method effectively addresses the common issues of existing approaches, which often select redundant similar samples or overly focus on boundary cases.
- 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 paper presents an innovative combination of representative-based and coverage-based methods for sample selection. By leveraging masks generated by the zero-shot foundation model SAM3 (via text prompts) to construct a diversified description space, the proposed method effectively addresses the common issues of existing approaches, which often select redundant similar samples or overly focus on boundary cases. 2.The overall organization of the manuscript is clear. The motivation is well-justified, and the proposed approach demonstrates a solid level of technical novelty. 3.The experimental results indicate that the proposed method achieves strong, state-of-the-art performance in terms of the Dice Similarity Coefficient (DSC) compared to existing baselines. Furthermore, the authors provide relatively comprehensive ablation studies to validate their network design choices.
- 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 scale of the experiments is somewhat limited. The evaluation is conducted on a relatively small number of datasets, which significantly weakens the persuasiveness regarding the generalizability and robustness of the proposed method. To make the claims more convincing, the authors are encouraged to evaluate their approach on more diverse and larger-scale datasets. 2.The manuscript lacks a thorough discussion on the scenarios where the proposed method underperforms. Specifically, the quantitative results show a disadvantage compared to some baseline methods under the Average Symmetric Surface Distance (ASSD) metric, but the authors fail to provide any insights or explanations for this. A detailed analysis of this limitation is necessary.
- 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?
The primary factors driving this positive score are the paper’s innovative approach to sample selection and its strong empirical performance. The motivation is well-justified, and the impressive improvements on the DSC metric clearly demonstrate the value of the proposed method. While there are a few areas for improvement, e.g., the somewhat limited scale of the evaluated datasets and the lack of explanation for the underperformance on the ASSD metric. These issues do not diminish the overall technical novelty and contribution of the work. Addressing these points will further solidify their claims.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
I recommend Accept because the paper presents a technically novel and well-motivated sample selection strategy. Overall, the work is solid, useful, and sufficiently convincing for acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes a dual-space cold-start active learning framework for medical image segmentation that uses SAM3-generated pseudo-masks as priors. The method combines feature-space typicality with mask-space morphological diversity to select informative and non-redundant samples under very small annotation budgets.
- 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 and practical problem in medical image segmentation, namely cold-start active learning under limited annotation budgets. The proposed dual-space design is intuitive and technically reasonable, combining representative sampling in feature space with morphology-aware diversity in pseudo-mask space. The experiments cover two public datasets, include multiple baselines and an ablation study, and show clear gains in Dice score under low-budget settings.
- 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 does not compare with the most relevant recent SAM-based CSAL baseline, especially SUGFW [15], which weakens the claim of state-of-the-art performance.
The methodological novelty appears moderate, since the main design combines representative sampling and diversity sampling ideas that are already central in prior methods such as TypiClust [6], BAL [9], and Core-set [19].
The empirical gains are more convincing in Dice than in boundary quality: ASSD is not consistently improved and is worse than several baselines in some settings.
Reproducibility is limited because several key details are unclear, including SAM3 prompt construction, the exact quality-score computation, feature extraction settings, and how the two spaces are normalized and combined.
- 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?
I appreciate the practical motivation and the generally solid experimental effort. However, I am not yet convinced by the current evidence because the comparison to the closest related method is missing, the gains are not consistent across all evaluation metrics, and several key implementation details are unclear. For these reasons, I view the paper as promising but currently below the acceptance threshold.
- 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 Cold-start Active Learning (CSAL) framework to mitigate selection redundancy and boundary outlier bias from both the feature space and the mask space, leveraging SAM3.In the feature space, a Centroid Affinity Scoring strategy is introduced, along with an adaptive Cluster-Balanced Penalty to reduce redundant queries dominated by major clusters. In the mask space, a Morphological Diversity exploration module with Quality Estimation is designed to capture diverse structural patterns. Experimental results on two public benchmarks demonstrate that the proposed method achieves state-of-the-art performance.
- 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 paper is generally well-written, with a clear structure and coherent presentation. The figures effectively illustrate the key ideas of the proposed framework.
2.The proposed modules are well-motivated, and the overall framework is technically sound. The dual-space design is conceptually consistent, and the Cluster-Balanced Penalty provides a reasonable mechanism to alleviate redundancy in the Centroid Affinity Scoring process.
3.The method achieves state-of-the-art performance on two public datasets, demonstrating its empirical effectiveness.
- 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.It would strengthen the paper to include ablation studies across different versions of SAM models, to better understand the robustness and generalizability of the proposed framework.
2.Additional visualizations illustrating the effect of the Cluster-Balanced Penalty in the feature space would improve interpretability and provide more intuitive insights into its contribution.
- 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?
Please see the strength part
- 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 for recognizing the motivation (R1/R2), design (R1/R2/R3), and performance (R1/R2/R3) of our method. We address the main concerns below:
*Novelty(R2) While representative sampling is a general CSAL principle, CSAL methods differ in definition of representativeness. Unlike existing methods[6,9,19] purely define it based on global image feature we further introduce object morphology-aware representativeness to encourage size/shape diversity based on SAM3 text-prompted pseudo-masks (Sec.2.1). We also introduce CBP to obtain a better feature-based representative score, and combine them in a unified score (Sec.2.3) for better sample selection.
*SUGFW comparison(R2) The source code of SUGFW was not actively maintained. We attempted to reproduce SUGFW under our segmentation backbone, but due to some unresolved issues in data preprocessing and hyper-parameter tuning, we failed to get results. We will contact the first author of SUGFW for help and encourage the author to fix the issues.
*ASSD analysis(R1/R2) Compared with Dice, ASSD is highly sensitive to small distant false-positive regions, which can disproportionately increase the surface distance. To diagnose the higher ASSD in the 15-sample PROMISE12 setting (Tab.1), we applied largest-connected-component post-processing as an auxiliary analysis. ASSD decreases from 6.15 mm to 2.83 mm, suggesting that the higher raw ASSD mainly comes from scattered disconnected false positives rather than poor segmentation of the main prostate region. This is consistent with our morphology-aware acquisition, which improves Dice by prioritizing structurally diverse foreground samples, but may yield more aggressive foreground predictions under limited budgets. As the budget increases, our method achieves the best ASSD with 100 samples. We will briefly explain this.
*Reproducibility(R2/MR) Following MedSAM3[13] and standard prompt template of SAM3, we prompt SAM3 with “skin lesion” for ISIC 2018 and “prostate” for PROMISE12 for reproducibility. For feature extraction, we use the penultimate layer of the SAM3 image encoder and spatial pool it for CAS. As described in Sec 2.2, the quality score is derived from SAM3’s final image-text cross-attention map[10] by projecting their attention to the spatial grid, taking the maximum token probability at each position, and averaging spatially. For normalization, centroid-affinity scores and morphological descriptors are min-max normalized over candidates, and mask-diversity distances are dynamically normalized at each greedy selection step before being combined. We will release the code for reproducibility.
*Dataset scale and SAM-variant robustness(R1/R3) Small sample size is an important and common challenge in medical image analysis, which motivates adapting pre-trained models like SAM/SAM3 to downstream medical image tasks. This work shows a cold start active adaptation of SAM3 for two medical image segmentation tasks with smaller dataset size than natural images. However, too small dataset will reduce the reliability of results, so we chose Prostate12 (100 cases) and ISIC 2018 (2594 cases) that are moderate to ensure reliable conclusion. SAM1/SAM2 were not included since their target-specific masks require point/box prompts, which conflicts with CSAL setting where no mask is provided beforehand. We use SAM3 because its text-prompted, concept-driven zero-shot inference enable obtaining pseudo-masks without additional annotation cost.
*Visualization of the CBP(R3) Due to space limits and the high-dimensional feature, a 2D visualization is difficult to include in the current manuscript. We therefore support this module through detailed ablation studies, where adding the penalty consistently improves performance. We will explore toy-data visualizations in an extended version to show how CBP avoids over-selection from dominant clusters.
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 paper is borderline-positive (scores 3–5), with reviewers recognizing a practical and well-motivated dual-space active learning framework, clear presentation, and strong Dice improvements. The rebuttal should mainly address the missing comparison to the closest recent SAM-based CSAL baseline, clarify the novelty beyond combining representative and diversity sampling, explain the weaker ASSD/boundary results, and improve reproducibility by detailing SAM3 prompts, quality scoring, feature extraction, and space normalization/combination. It would also help to comment on robustness across SAM variants and dataset scale. Overall, the concerns are focused and addressable, so a clear rebuttal could strengthen the case for acceptance.
- 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.
Prior to the rebuttal, the paper received a single negative rating (WR). However, this reviewer did not provide any follow-up comments after the authors’ rebuttal. In contrast, the rebuttal successfully addressed the concerns raised by the other reviewers and led to a more favorable overall assessment. Based on the reviewers’ feedback and the authors’ responses, the AC recommends 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.
This submission somehow received mixed scores, with reviewers mainly questioning the actual novelty, and clarifications on results and reproducibility. After carefully reading authors rebuttal, I believe that these concerns were satisfactorily addressed, and thus recommend the acceptance of this paper.
I encourage the authors to incorporate these clarifications and improvements into the camera‑ready version
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.
I recommend rejection. The paper addresses an important problem, namely cold-start active learning for medical image segmentation, but the current submission has several significant weaknesses.
First, the methodological novelty is limited. The proposed framework mainly combines existing ideas: centroid/typicality-based sampling, cluster balancing, farthest-point-style diversity sampling, and pseudo-mask descriptors from a foundation model. The paper does not clearly establish why this combination is technically new beyond an engineering aggregation of known components.
Second, the use of SAM3 is insufficiently justified and under-specified. The method relies heavily on SAM3 features, pseudo-masks, and cross-attention-based quality scores, but the paper does not clearly define the prompting strategy, the mask-generation protocol, or why the proposed attention-based uncertainty is a reliable proxy for medical segmentation quality. This is particularly problematic because the self-training baseline shows highly unstable pseudo-mask quality, especially on ISIC 2018. Third, there are important technical ambiguities in the acquisition function. The figure includes a weighting parameter between the image-space and mask-space scores, but the final equation omits it. The normalization of the mask distance, initialization when no sample has been selected, the choice of the number of clusters, and the computation of the quality score are not sufficiently specified. These details are critical for reproducibility.
Fourth, the experimental evidence does not strongly support the claimed superiority. Many reported improvements are small relative to the very large standard deviations. For example, on ISIC 2018 with 100 samples, the DSC gain over strong baselines is marginal, while the ASSD is worse than several methods. On PROMISE12 with 15 samples, the proposed method has the best DSC but a relatively poor ASSD. Therefore, the claim of consistent and substantial improvement is overstated.
Finally, the evaluation scope is limited. Only two datasets are used, and the annotation unit is not fully clear for PROMISE12, where 2D slices are selected but 3D volumetric evaluation is reported. The paper would also benefit from stronger statistical testing, clearer baseline implementation details, and broader validation across more medical modalities.