List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised medical image segmentation (SSMIS) has garnered significant attention for its ability to reduce radiologists’ manual annotation burden. However, inadequate distinction between different categories in the feature space can lead to difficulty in distinguishing visually similar classes. Vision–language models like CLIP excel at disambiguating such categories through rich semantic representations, but their coarse‑grained embeddings struggle to capture fine structural details in complex medical scans. To address these challenges, we propose a CLIP-based SSMIS framework, which comprises three key components: unified representation matching (URM), prototype alignment (PA) and embedding fusion CLIP (EFCLIP). Specifically, URM aligns spatial feature representations between teacher and student networks to mitigate the impact of high-uncertainty voxels and ambiguous boundaries. Building on this, PA selects high-confidence voxels based on voxel-level uncertainty to construct robust class prototypes. Furthermore, EFCLIP integrates CLIP’s text embeddings with voxel-level visual features, fusing global semantic priors with local spatial structure to enhance the discriminative ability for visually similar voxels categorize. Extensive experiments on one public and one in-house medical image datasets demonstrate that PAVL consistently outperforms state-of-the-art SSMIS methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0625_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{KuaJun_PAVL_MICCAI2026,
author = { Kuang, Junchang AND Chen, Jiale AND Feng, Siyang AND Liu, Zaiyi AND Lin, Huan AND Lan, Rushi AND Pan, Xipeng},
title = { { PAVL: Enhancing Feature Prototype Alignment with Vision–Language Model for Semi-supervised Medical Image Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes PAVL, a CLIP-based semi-supervised medical image segmentation framework comprising three components: Unified Representation Matching (URM), Prototype Alignment, and Embedding Fusion CLIP. The framework is evaluated on BraTS2019 and an in-house lung adenocarcinoma CT dataset under 10% and 20% labeled data settings, reporting consistent improvements over seven SSMIS 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 PA module’s formulation of voxel embeddings as multivariate Gaussian distributions, combined with inverse-variance weighting for prototype construction and Wasserstein-2 distance for alignment, is theoretically well-motivated. 2.Table 2 offers a well-structured ablation that goes beyond simple add/remove comparisons: EFCLIP is compared against LoRA fine-tuning and CLIP-only guidance; URM is compared against standard KD and SRD distillation; and the PA module’s Wasserstein-2 distance is compared against Euclidean distance. This design allows readers to evaluate the specific contribution of each methodological choice rather than just the module as a whole. 3.PAVL achieves statistically significant improvements (p < 0.05, two-sided paired t-test) on most metrics across both BraTS2019 and the in-house dataset under both 10% and 20% labeled 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.The central claim of PAVL is that CLIP’s semantic representations help distinguish visually similar classes in the feature space. However, both evaluation datasets involve binary segmentation tasks (tumor vs. background), where there are no multiple visually similar classes to disambiguate. Neither dataset provides a scenario that meaningfully tests inter-class confusion. The overall DSC improvement observed does not constitute evidence that the method specifically addresses visually similar class confusion. As currently presented, there is a logical gap between the stated motivation and the experimental evidence. 2.PAVL adds substantial components on top of the baseline. However, the paper provides no comparison of parameter counts, FLOPs, or training/inference time against any baseline. In semi-supervised learning evaluation, where unlabeled data can be better leveraged by larger models, this confound is particularly important to address. 3.Section 3.3 contains several sentences reporting Pancreas-CT results (“PAVL sets new state-of-the-art performance, achieving DSC scores of 75.79% and 81.62% with only 10% and 20% labeled data”), yet this dataset is neither described in Section 3.1 nor present in Table 1.The omission raises questions about why Pancreas-CT results were removed and whether selective reporting may be involved.
- 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
The authors are encouraged to: (1) either replace the binary segmentation benchmarks with a multi-class dataset (e. g. , multi-organ segmentation, lesions with similar appearance) to validate the visually similar class claim, or add targeted analysis of boundary-ambiguous regions using a dedicated metric; (2) provide a parameter count and FLOPs comparison table against all baselines; (3) reconcile the Pancreas-CT experimental results
- 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 proposes a methodologically coherent framework with a well-motivated probabilistic treatment of voxel uncertainty and a reasonable integration of CLIP semantics into SSMIS. The ablation study is more thorough than typical in the field. However, the core motivation—distinguishing visually similar classes—is not validated by the experimental setup. Second, the absence of any parameter or efficiency analysis leaves open the alternative explanation that gains stem from model capacity rather than design. And also some minor flaws are in the current form of this paper. These issues are addressable in revision, and the underlying ideas have merit, but the current submission requires substantial revision.
- 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
This paper proposes a novel SSMIS framework based on feature prototype alignment with vision–language (PAVL) to address class confusion in the feature space and CLIP’s limited ability to capture fine-grained spatial structures. PAVL employs the URM strategy to align spatial features between teacher and student models, utilizes PA module to construct more accurate class prototype distributions by voxel level uncertainty, and integrates CLIP’s global semantic prior knowledge with voxel-level local spatial structures. Experiments conducted on the public dataset BraTS2019 and internal medical image datasets demonstrate that PAVL outperforms certain SSMIS methods.
- 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 integrates three components (URM, PA, EFCLIP) into a unified SSMIS framework, offering a relatively complete solution.
- URM explicitly calibrates spatial features on high-uncertainty voxels and fuzzy boundaries, concentrating model capacity on hard samples more effectively than global consistency regularization.
- PA constructs category prototypes by selecting high-confidence voxels based on voxel-level uncertainty, reducing noise and ensuring stable prototype representation.
- EFCLIP uses element-wise product to fuse CLIP text embeddings with softmax probability maps, potentially improving discriminability for visually similar voxel classes.
- 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 Text-SemiSeg (MICCAI 2025), which shares the exact task (CLIP-enhanced SSMIS) and similar modular design, and reports competitive results on the same datasets. Link: https://doi.org/10.1007/978-3-032-04971-1_57
- PA is highly similar to UPCoL (MICCAI 2023), PCCS (2025), and EPL (2024), which also use uncertainty masks to build/refine prototypes. The paper provides no direct comparison or theoretical differentiation. Links: https://doi.org/10.1007/978-3-031-43901-8_63; https://arxiv.org/abs/2502.06650; https://arxiv.org/abs/2404.06181
- The article does not clarify the advantages and disadvantages of element-wise product in EFCLIP compared to more commonly used cross-attention and additive fusion methods.
- 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 proposed framework shows acceptable novelty in system integration and the lightweight EFCLIP design. However, the experimental validation is insufficient to support the claimed superiority. The omission of the most relevant baseline (Text-SemiSeg, MICCAI 2025) and the lack of direct comparison with existing uncertainty+prototype methods (UPCoL, PCCS, EPL) critically undermine the credibility of the results.
- 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
This paper addresses two core challenges in semi-supervised medical image segmentation: poor inter-class distinction in feature space and the failure of CLIP’s coarse-grained embeddings to capture fine structural details in medical scans. It proposes PAVL, a CLIP-based framework composed of unified representation matching, prototype alignment and embedding fusion CLIP modules.
- 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 well-written with clear logical structure. The narrative flows smoothly from problem statement to method design and experimental validation. 2.It targets common challenges in semi-supervised medical image segmentation and proposes a three-module framework with a complete technical route. 3.The experiments are conducted on both public and in-house datasets under 10% and 20% annotation ratios, providing quantitative and qualitative results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Major Weaknesses
1.Core modules URM and EFCLIP lack substantial novelty. URM is essentially the classic teacher-student feature-level consistency constraint in semi-supervised learning, only adding simple convolutional modules for adaptation, no substantial methodological innovation. EFCLIP’s text-visual fusion is highly consistent with prior text-based methods, only performing basic semantic fusion, no differentiated innovation.
2.PA module has limited novelty, only formal substitutions to classic methods. Its core logic remains selecting high-confidence samples to build class prototypes, only replacing conventional uncertainty measurement with feature variance, without breaking existing prototype learning-based semi-supervised segmentation frameworks, negligible incremental innovation.
3.Comparative experiments. Most methods in Table 1 are traditional semi-supervised methods, excluding latest foundation model-based SOTAs of the same type, failing to fairly verify the proposed method’s real performance advantage.
4.Ablation study. Table 2 lacks the key control group removing EFCLIP and retaining only URM and PA, making it impossible to separate gains from CLIP pre-training priors and the proposed modules’ contributions.
Minor Weaknesses
1.Method is only validated on two binary 3D segmentation tasks, not covering multi-class segmentation.
2.No limitations of the method are discussed, including computational overhead, inference speed, parameter count and stability.
- 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?
The overall score is mainly determined by the limited novelty of the proposed modules, unfair comparison with recent SOTA methods, and flawed ablation design, which together weaken the methodological contribution and experimental reliability of this work.
- 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
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.
This work exhibits strong practical value by systematically utilizing CLIP to resolve inter-class ambiguity and refine boundaries in semi-supervised feature spaces. Reviewers commend the mathematical rigor of the Prototype Alignment module, which effectively models voxel-level uncertainty via Wasserstein-2 distance to construct robust class prototypes.
