List of Papers Browse by Subject Areas Author List
Abstract
In-context learning (ICL) has recently emerged as a promising paradigm for medical image segmentation, enabling frozen pre-trained models to adapt to new tasks using only a few labeled demonstrations (context). However, the effectiveness of ICL critically depends on the quality and representativeness of the provided context set. Redundant, noisy, or suboptimal demonstrations not only degrade segmentation accuracy but also increase computational overhead during inference. In this work, we propose Context Distillation via Masked Image Modeling (CD-MIM), a framework that reformulates context construction as a task of learning compact and task-specific anatomical priors. Instead of relying solely on raw clinical samples, we treat the context as a learnable entity and distill it into a small set of optimized templates. Through a curriculum-driven masking strategy and a meta-learning-based template optimization process, CD-MIM encourages the model to encode essential structural dependencies and cross-patient anatomical variations into condensed context. The resulting distilled context can either replace or complement raw demonstrations, reducing redundancy while improving segmentation robustness. Extensive experiments on seven medical image segmentation tasks and three representative visual in-context learning backbones demonstrate consistent performance gains and improved efficiency. Our findings suggest that explicitly learning distilled context priors provides a scalable and principled solution for stabilizing and enhancing ICL-based medical image segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3064_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{YanGuo_Learning_MICCAI2026,
author = { Yang, Guoqing AND Hu, Shishuai AND Xia, Yong},
title = { { Learning Task-Specific Anatomical Priors via Context Distillation for In-Context Medical Image 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
This paper proposes CD-MIM, a context distillation framework for in-context medical image segmentation. Instead of selecting raw support examples from a candidate pool, the method treats the context itself as a learnable object and distills it into compact, task-specific templates that encode anatomical priors. The framework combines a curriculum-driven masking strategy with a meta-learning-style optimization process, with the goal of encouraging the distilled templates to capture structural dependencies and cross-patient anatomical variation while suppressing redundancy and noise. The method is evaluated on seven segmentation tasks across multiple imaging modalities using three representative in-context segmentation backbones, and the results suggest that distilled context can improve performance over several selection-based baselines.
- 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 dependence of VICL-style segmentation on context quality is a real limitation, especially when the candidate pool grows large. The idea of learning compact context priors instead of repeatedly selecting raw demonstrations is well motivated.
2.Reframing context construction as a distillation problem is a meaningful conceptual contribution. This is a more proactive strategy than standard retrieval or ranking-based context selection.
3.The paper evaluates seven tasks across five modalities and tests three different VICL backbones.
- 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.All experiments use context size = 1, which makes the problem easier to analyze, but also limits the practical scope of the conclusions. Since many in-context segmentation frameworks can use multiple demonstrations, it would be important to know whether the method still helps when the support size is larger.
2.The paper includes an ablation comparing raw vs distilled templates, but there is little analysis of the individual design choices, such as the effect of the masking schedule, the choice of masking ratio bounds, the benefit of the curriculum itself versus fixed masking, or the contribution of the STE-based mask optimization.
3.The method is clearly effective for UniverSeg and Tyche-TS, but on SegGPT, the average score is below Top-K and VPR in Table 1.This weakens the claim of consistently superior performance across representative VICL backbones.
4.A central motivation is reducing redundancy and computational overhead, but the paper does not report actual runtime, memory, or retrieval-cost comparisons.
5.Distilling templates separately within each task setting appears useful, but it is less clear how well this scales to more realistic deployment scenarios where task boundaries may be ambiguous or new anatomical targets arise.
6.Some early but important medical dataset distillation methods [1-3] are not cited.
[1] Soft-Label Anonymous Gastric X-ray Image Distillation [2] Compressed Gastric Image Generation Based on Soft-Label Dataset Distillation for Medical Data Sharing [3] Dataset Distillation for Medical Dataset Sharing
- 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.
(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?
On the positive side, the paper addresses a meaningful problem in in-context medical image segmentation, and the idea of distilling task-specific anatomical priors into compact context templates is novel and well motivated. The results on UniverSeg and Tyche-TS are encouraging, and the overall framework is clearly presented. At the same time, I am not fully convinced that the current evidence is strong enough for acceptance. The evaluation is centered on the context-size-one setting, which limits the scope of the conclusions. The ablation study is also relatively light, so it remains unclear which parts of the method are most responsible for the improvement. In addition, the results are less convincing on SegGPT, and the efficiency benefits discussed in the paper are not quantitatively demonstrated. Overall, I find the paper promising, but still a bit short of being fully convincing in its current form. A strong rebuttal and clearer analysis of these points could improve my assessment.
- Reviewer confidence
Very confident (4)
- [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.
Based on the rebuttal, I find that most of my major concerns have been adequately addressed. There are still a small number of remaining issues, but overall, I am willing to recommend Accept.
Review #2
- Please describe the contribution of the paper
The paper proposes a new framework, Context Distillation via Masked Image Modeling (CD-MIM), for in-context medical image segmentation. The main idea is to move away from using raw support images as context and instead learn compact, task-specific context templates that capture important anatomical information. Unlike existing methods that focus on selecting or matching the best support examples, this work directly learns and optimizes the context itself. These templates are trained to encode meaningful structural and shape-related information while reducing noise and redundancy present in real images. To ensure that the templates learn true anatomical structure rather than superficial textures, the authors introduce a curriculum-driven masking strategy. During training, increasing portions of the support images are masked, and the templates are used to fill in the missing regions. As masking becomes more aggressive, the templates are forced to capture strong structural priors and cross-patient variations, thereby becoming more informative and robust. Additionally, the templates are optimized using a meta-learning setup, in which they are trained to generalize across different samples and tasks, thereby improving their effectiveness in real in-context segmentation scenarios. Overall, the key contribution is a learned, compact, and robust context representation that replaces raw examples, reduces redundancy, and leads to more stable, accurate segmentation performance across multiple datasets and model backbones.
- 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 has several notable strengths. 1.The paper introduces a fundamentally new perspective by treating the context not as a fixed set of support examples but as a learnable and optimizable entity. This shift from context selection to context representation learning is novel and interesting, as it allows the model to directly encode task-relevant anatomical priors while reducing redundancy and noise inherent in raw medical images. 2.The use of masked image modeling (MIM) is not merely for self-supervision, but is cleverly adapted as a distillation constraint. By forcing templates to reconstruct masked regions, the method encourages learning of global structure and spatial dependencies, rather than relying on superficial texture cues. This is particularly important in medical imaging, where structural consistency is critical. 3.The introduction of a curriculum-based masking scheduler is a strong design choice. By gradually increasing the masking ratio, the model follows an easy-to-hard learning paradigm, first capturing coarse anatomical structure and then finer details. This progressive constraint ensures that the learned templates become highly informative and robust under limited information conditions. 4.The use of a meta-learning setup to optimize the templates is well-motivated. By simulating query-support splits during training, the templates are encouraged to generalize across different samples and patients, making them more effective in real-world in-context segmentation scenarios. 5.By replacing multiple raw support images with a small set of learned templates, the method significantly reduces redundancy and memory overhead. At the same time, it improves robustness by filtering out noise and irrelevant variations, which is a practical advantage for deployment. 6.The method is thoroughly validated on seven segmentation tasks using three visual in-context learning backbones, demonstrating its consistency and robustness across different datasets, modalities, and architectures. This extensive evaluation strengthens the credibility and generalizability of the proposed approach. 7.The approach is well-aligned with the needs of medical image analysis, as it explicitly encourages the learning of anatomical structure and cross-patient morphological variations. This makes the method both practically relevant and scientifically meaningful for clinical applications.
- 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 has some shortcomings 1.The paper is written in a highly technical manner, making it difficult for readers to follow. Key concepts (e.g., templates, masking strategy, optimization flow) lack intuitive explanations or illustrative examples, which affects overall readability and understanding. 2.The evaluation appears to rely primarily on a limited set of metrics, making it difficult to draw comprehensive conclusions about performance. Incorporating additional metrics (e.g., boundary-based metrics, calibration, robustness measures) would provide a more holistic assessment of segmentation quality. 3.The contribution involves multiple components (masking scheduler, MIM constraint, meta-learning), but the paper provides limited, detailed ablation studies to isolate the impact of each component. 4.While the method reduces redundancy at inference, the training procedure involves masking, meta-learning, and template optimization, which may introduce additional computational complexity. This aspect is not clearly analyzed. 5.The learned templates are central to the method, but their interpretability is not explored. Visualizing or analyzing what the templates capture would strengthen the paper’s insights.
- 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.
(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 presents a novel and well-motivated approach by introducing context distillation for in-context medical image segmentation, shifting from raw support selection to learnable context representations. The integration of masked image modeling as a distillation constraint and the curriculum-driven masking strategy are particularly interesting design choices that encourage the learning of meaningful anatomical structure rather than superficial features. Additionally, the use of meta-learning for template optimization, along with a comprehensive evaluation across seven segmentation tasks and three backbones, demonstrates the robustness and generalizability of the approach. However, the overall impact is somewhat limited by concerns about clarity and evaluation. The paper is written in a highly technical manner, making it difficult to grasp key ideas without significant effort, which affects accessibility. Furthermore, the evaluation relies on a limited set of metrics, making it harder to assess the quality and clinical relevance of the results fully. There is also limited discussion on computational overhead and interpretability of the learned templates. Overall, while the paper introduces a promising and novel direction with strong experimental validation, improvements in clarity, broader evaluation, and deeper analysis would strengthen its contribution and practical impact.
- 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.
The authors have addressed all the raised concerns and I am satisfied with the responses. In particular: 1.The authors clarified the role of templates, masking strategies, and optimization flow, improving methodological clarity. 2.The computational efficiency argument is compelling since the distillation process is performed only once offline, while inference avoids expensive per-sample retrieval. 3.Although some limitations remain, such as performance gaps on challenging lesion segmentation tasks and limited analysis of unseen task boundaries, the authors clearly acknowledged these points and positioned them as future work rather than unsupported claims.
Review #3
- Please describe the contribution of the paper
This paper proposes Context Distillation via Masked Image Modeling (CD-MIM), which introduces context distillation into medical vision context-aware segmentation for the first time, redefining context construction as the learning of task-specific compact anatomical priors. We designed a Curriculum-driven Masking Scheduler coupled with meta-learning template optimization, which forces the model to encode cross-patient structural dependencies and morphological priors into learnable templates under increasingly sparse visible information conditions. We validated the method’s consistency and robustness across 7 segmentation tasks (5 imaging modalities) and 3 representative VICL backbones (UniverSeg, Tyche-TS, SegGPT).
- 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 Distilled Template proposed in this paper suppresses redundant information and out-of-distribution noise in the original clinical samples (Fig. 2), demonstrating better boundary accuracy and structural consistency on challenging queries. The distilled context can replace the original demonstrations, compressing the supporting information into a small number of optimized templates, thereby reducing the context length and computational overhead during inference and improving inference efficiency. Consistent improvements are achieved across both cross-architecture (UniverSeg/Tyche-TS vs. ViT-based SegGPT) and cross-modal (fundus, X-ray, MRI, ultrasound, etc.) settings, demonstrating a degree of generalizability.
- 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 performance improvement achieved by CD-MIM on SegGPT (+9.32%) is significantly smaller than that of medical-specific backbones. The authors attribute this to the fact that SegGPT uses a ViT pre-trained on natural images, which is less sensitive to medical anatomical prior knowledge. Does this imply that the performance gains from CD-MIM are highly correlated with the domain-specific suitability of the backbone? Template initialization currently employs stochastic selection; although the authors mention that diversity strategies could be used to enhance it, they do not elaborate. Furthermore, distillation requires an additional 200 epochs of training, increasing computational costs prior to deployment. The proposed method performs poorly in absolute terms on certain tasks. For example, on the UniverSeg LGG task, the DSC after distillation is only 20.95%, and on GLS it is merely 47.78%; the overall average performance also falls short of the best results. This indicates that there are still significant shortcomings when it comes to segmenting certain high-difficulty lesions. The ablation experiments are too general and provide almost no specific analysis of the method’s effectiveness.
- 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?
Writing logic and experimental results.
- Reviewer confidence
Very confident (4)
- [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.
The author has provided additional details regarding the ablation study, which demonstrate the validity of the proposed method.
Review #4
- Please describe the contribution of the paper
The paper introduces a context distillation framework for in-context medical image segmentation, where a small set of learnable templates is optimized to replace raw demonstrations, enabling more compact and task-specific conditioning of frozen segmentation models.
- 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 presents a clear and meaningful reformulation of in-context medical image segmentation by shifting from context selection to context distillation, i.e., learning compact task-specific templates rather than selecting raw demonstrations from a large candidate pool. This is an interesting extension of dataset distillation to the inference-time conditioning setting of VICL. The proposed use of masked image modeling as a distillation constraint is also well motivated, as it encourages the templates to encode structural information under incomplete support input. In addition, the curriculum-driven masking strategy is simple and reasonable. The method is evaluated on three VICL backbones and seven segmentation tasks across multiple modalities, which strengthens the empirical scope of the work. The gains on medical-specific backbones, especially UniverSeg and Tyche-TS, are substantial and suggest practical value in low-context segmentation 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.
1.Several claims are somewhat stronger than what the current experiments support. In particular, the paper suggests broadly consistent improvements across backbones, but Table 1 shows that CD-MIM is not the best method on average for SegGPT. The corresponding claims should therefore be stated more carefully. 2.The paper emphasizes improved efficiency and reduced redundancy, but does not provide direct empirical evidence for these points, such as runtime, memory usage, or scalability as the candidate pool grows. This makes the practical advantage less convincing than it could be. 3.The claim that the distilled templates capture task-specific anatomical priors is only partially supported. At present, this conclusion relies mainly on quantitative improvements and limited qualitative visualization, without deeper analysis of what is actually encoded in the learned templates. 4.The experimental setting is somewhat narrow. All results are reported with context size 1, so it remains unclear how well the method generalizes to broader in-context settings.
- 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 makes a meaningful contribution by recasting in-context medical image segmentation as a context distillation problem, replacing raw demonstrations with compact task-specific templates. This is a well-motivated idea. Several issues keep me from rating the paper more highly. Some of the claims are broader than what the results fully support, especially since the improvements are not consistent across all backbones. The paper also highlights efficiency and reduced redundancy, but does not provide direct evidence such as runtime, memory, or scaling analysis. In addition, the claim that the learned templates capture task-specific anatomical priors is only partially supported by the current experiments. Finally, the evaluation setting is somewhat limited, as all experiments use context size 1 and the ablation study does not fully separate the effects of the main design choices.
- 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.
N/A
Author Feedback
We sincerely thank the reviewers for their comments and recognition of our work.
R1Q1/R4Q4: Context size. We fixed the context size to 1 mainly due to space limitations and to isolate the effect of context distillation. We further evaluated K=2 with diversity-based initialization. CD-MIM improved over raw support samples, achieving 67.10% vs. 56.06% DSC with UniverSeg and 72.64% vs. 65.91% with Tyche-TS, demonstrating effectiveness with multiple demonstrations.
R1Q2/R2Q3/R3Q5: Ablation studies. Due to space limitations, additional ablations were omitted. On UniverSeg, removing the masking schedule reduced DSC to 31.29%, validating the MIM strategy. Fixed masking ratios (25%/50%) achieved 48.15%/59.40%, both worse than the dynamic curriculum. Different masking bounds ([0.1,0.5]/[0.4,0.6]/[0.5,0.9]) achieved 56.75%/62.73%/67.71%, highlighting the importance of proper bounds. STE is also necessary for mask template optimization.
R1Q3/R4Q1: Performance using SegGPT. Although CD-MIM is slightly below Top-K on SegGPT, it still improves the raw baseline by +9.32%, demonstrating its plug-and-play effectiveness. We believe the remaining gap is mainly related to unified hyperparameter settings across different backbones rather than architecture-specific tuning, and will further investigate this in future work.
R1Q4/R2Q4/R3Q3/R4Q2: Computational costs. The distillation cost is incurred only once before deployment. During inference, CD-MIM directly uses the distilled template and avoids the per-sample retrieval required by methods such as VPR. For example, VPR required 479 seconds for 6,519 test samples, while CD-MIM eliminates this stage entirely.
R1Q5: Scalability to realistic scenarios. We agree that handling ambiguous task boundaries and unseen anatomical targets is important for realistic VICL deployment. While beyond the scope of this work, we believe context distillation provides a promising direction and will further investigate this problem in future work.
R1Q6: Related works on dataset distillation. We will include and discuss these related works in the final version.
R2Q1: Concept explanations. The “template” is a learnable prompt distilled from the context pool, the “masking strategy” follows masked image modeling for gradual template optimization, and the “optimization flow” is formulated as a meta-learning process.
R2Q2: Additional metrics. We believe the DSC improvements across seven segmentation tasks demonstrate the effectiveness of CD-MIM. We also agree that additional metrics could provide a more comprehensive evaluation and will include them in future work.
R2Q5/R4Q3: Visualization and analysis. As shown in Fig. 2, the distilled template captures shared spatial priors across diverse context samples rather than memorizing the initialization sample. We will further improve visualization and interpretability analysis in the final version.
R3Q1: Domain-specific suitability. The purpose of evaluating multiple VICL backbones, including UniverSeg, Tyche-TS, and SegGPT, is to validate the generality of CD-MIM across architectures. While absolute gains may depend on training data and backbone design, analyzing these intrinsic differences is beyond this paper’s scope. Importantly, CD-MIM improves performance across all evaluated backbones, demonstrating its effectiveness.
R3Q2: Template initialization strategy. We additionally explored diversity-based initialization for K=2, improving DSC from 56.06% to 67.10% on UniverSeg and from 65.91% to 72.64% on Tyche-TS, showing that CD-MIM can effectively integrate principled initialization strategies.
R3Q4: Performance on high-difficulty lesions. Our work focuses on context distillation, while challenging lesion segmentation also depends heavily on the intrinsic capability of the VICL backbone. Thus, improvements from context distillation alone may be limited on particularly difficult tasks. We will further investigate this direction in future work.
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 interesting overall and presents a promising idea, but reviewer concerns remain regarding limited experimental settings (e.g., context size), insufficient ablation and analysis of key components, inconsistent performance across backbones, and lack of quantitative evidence for efficiency and interpretability claims; these issues, along with some missing details, can likely be addressed during the rebuttal.
- 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.
I recommend Accept after rebuttal. The rebuttal adequately addressed the main concerns by clarifying the template design, masking strategy, optimization flow, computational efficiency, and ablation evidence, while appropriately acknowledging remaining limitations such as challenging lesion segmentation cases and unseen task boundaries as future work rather than overstated claims.
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.
While this submission received mixed scores during the initial stage, the rebuttal positively addressed the raised concerns, particularly regarding on the ablation studies, computational costs and further clarifications/discussions.
I side with the four reviewers and recommend Accept for this work.
Meta-review #3
- 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 novel context-distillation framework for in-context medical image segmentation, replacing raw support examples with compact task-specific anatomical templates. Although reviewers initially raised concerns about methodology and experiment details, the rebuttal addressed these points. Overall, all reviewers updated or maintained their final recommendation as Accept, supporting acceptance.