Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Semi-supervised medical image segmentation alleviates the burden of costly expert annotation. Textual information provides complementary guidance to enhance visual feature learning under limited supervision. However, existing vision-language methods operate in Euclidean spaces, which fail to capture inherent tree-like anatomical hierarchies despite the benefits of textual knowledge, leading to geometric distortion. In this paper, we propose \textbf{HyperSemi}, a hyperbolic vision-language interaction network for semi-supervised medical image segmentation. Our framework is specifically designed to accommodate anatomical hierarchies through geometry-aware cross-modal learning. Specifically, we present a Hyperbolic Semantic Embedding (HSE) module, which projects both visual features and text embeddings into the Poincaré ball to enforce hierarchical semantic alignment, effectively mitigating geometric distortion. To enable geometry-consistent cross-modal interaction, we further propose a Geometry-aware Semantic Integration (GSI) module that leverages hyperbolic distance weighting and tangent-space aggregation to adaptively modulate encoder features. Finally, to mitigate confirmation bias introduced by pseudo-labels, we propose an Uncertainty-aware Region Mixing (URM) strategy that selectively identifies high-uncertainty regions for mixed supervision, improving model robustness. Extensive experiments on three benchmark datasets demonstrate that our model consistently outperforms state-of-the-art semi-supervised segmentation approaches.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2184_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/taozh2017/HyperSemi

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HeQiu_Hyperbolic_MICCAI2026,
        author = { He, Qiuchi AND Jin, Shaocheng AND Wang, Rui AND Xu, Tianyang AND Wu, Xiao-Jun AND Zhou, Tao},
        title = { { Hyperbolic Vision-Language Interaction 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

    The paper introduced HyperSemi presents a novel geometric perspective for semi-supervised medical image segmentation, and the application of hyperbolic space to model anatomical hierarchies is a well-motivated idea. The proposed modules takes the anatomical hierarchies and aligns cross-modal features, in this way the discriminative representation is preserved. The multi-dataset evaluation and thorough ablation study are also presented as a support of idea.

  • 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.

    Using hyperbolic geometry to capture the inherent hierarchical structure of medical anatomy is attractive and new that fits to the problem, probably a more systematic and mathematically sound way of dealing the inhomogeneities. The motivation that Euclidean space introduces distortion when embedding tree-like anatomical concepts is well connected to move beyond. Evaluated on diverse datasets, which makes it good in terms of generalizability. The proposed modules can be plugged into existing frameworks, therefore hyperbolic alignment idea has broader applicability beyond the specific HyperSemi framework.

    The mathematical formulation of the Poincare ball operations, hyperbolic distance metric, exponential and logarithmic maps, and the resulting modules is presented in a clear way.

  • 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 central motivation of this paper is that hyperbolic space better captures anatomical hierarchies. However, no analysis is provided to show that the learned hyperbolic embeddings actually encode this hierarchy. There are no visualizations of the Poincare ball embeddings, no hierarchy probing experiments, and no qualitative analysis of the learned representations Probably at least add a visualization of learned embeddings.

    The curvature parameter c is a fundamental hyperparameter of the Poincare ball model that controls the geometry of the space and directly affects the capacity to represent hierarchies. The paper uses this parameter throughout all formulations but never states what value was used, whether it was fixed or learned, or whether the model is sensitive to its choice.

    CLIP is pre-trained on 2D natural image-text pairs, and applying it to 3D volumetric medical data introduces a significant domain gap. The paper acknowledges this but addresses it only through CoOp prompt tuning on the text side. No experiment isolates the effect of this domain gap or compares a 3D-aware text-visual model against the 2D CLIP baseline. The URM strategy selects high-uncertainty regions based on prediction discrepancy between dual networks and uses them for mixed supervision. However, similar uncertainty-driven region selection and mixing strategies have been proposed in prior work.

    The paper uses Context Optimization (CoOp) to learn continuous prompts for the text encoder. These learned prompts are a core part of the HSE module. However, the paper never examines or discusses what these learned prompts represent or whether they capture meaningful domain-specific knowledge.

    Under the 10% labeled setting, HyperSemi achieves a Dice score of 86.12 on BraTS-2019 compared to 85.30 for PICK and 85.27 for Text-SemiSeg. Under 20% labeled data, the improvement over Text-SemiSeg is 0.52 Dice points. The paper does not discuss why the gains are noticeably smaller on this dataset compared to LA and Pancreas.

    Table 3 is presented as a fairness analysis to show that incorporating textual cues is generally beneficial. However, this is performed on only two methods (MCF and BCP) and two datasets (LA and Pancreas). The comparison also does not include the full HyperSemi pipeline, only the HSE module.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    Add visualization of learned Poincare ball embeddings to validate the hierarchical representation claim, which is the central motivation of the paper. Report the curvature parameter c and include a sensitivity analysis over its values. Add a direct ablation comparing hyperbolic HSE against an equivalent Euclidean cross-modal alignment to isolate the geometric contribution. This is the most critical missing experiment. Extend the HSE plug-in analysis in Table 3 to more methods and datasets. Include statistical significance testing across multiple runs with different random seeds. Discuss and analyze the learned CoOp prompt tokens to provide interpretability for the text-guided learning component. Provide a clear discussion of the smaller gains on BraTS-2019 compared to LA and Pancreas.

  • 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?

    Good approach, sound concept implemented inside the framework, Justified with strong motivation against Euclidean methods. Sound experiments.

  • 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

    HyperSemi combines hyperbolic vision-language alignment, geometry-aware cross-modal fusion, and uncertainty-driven region mixing for semi-supervised medical image segmentation. Consistently outperforms 11 recent baselines on LA, Pancreas, and BraTS-2019 segmentation tasks.

  • 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 use of hyperbolic geometry for vision-language alignment is an interesting direction. Results are consistent across datasets and metrics (Dice, 95HD, ASD). Ablation shows each component contributes independently.

  • 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.

    Hyperbolic operations are applied to globally pooled vectors, not to 3D spatial representations. The geometric properties of the Poincaré ball are therefore never leveraged over the volumetric structure the paper is motivated by. There is no ablation comparing hyperbolic alignment to a standard Euclidean equivalent at the same feature level. Without this, the benefit of the hyperbolic projection over cosine or dot-product alignment cannot be assessed. The curvature hyper-parameter $c$ is not reported.

  • 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?

    Results are consistent across multiple datasets. However, the missing Euclidean baseline leaves the contribution of hyperbolic geometry unverified.

  • 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 is significant in that it formulates Vision-Language Alignment for medical image semi-supervised learning within a hyperbolic space.
    • Rather than introducing a completely new paradigm, this work represents a highly effective engineering contribution, achieving state-of-the-art performance by combining existing methods.
    • A notable strength is the clear role of each proposed module, and the synergistic relationships among them are well-validated through comprehensive experiments.
    • Finally, achieving such training efficiency and SOTA performance using a single RTX 3090 GPU demonstrates the high practical value of this research.
  • 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 work is meaningful in medical SSL because it explicitly reframes vision–language alignment in a hyperbolic space, which matches recent interest in structured multimodal representations while addressing the mismatch between flat Euclidean VL embeddings and hierarchical anatomy; this is a clear conceptual hook beyond “CLIP + segmentation.”
    • Rather than claiming a full paradigm shift, the paper is best read as efficient engineering: it connects established ingredients—semi-supervised dual networks, CLIP text conditioning, prompt learning (CoOp [32]), and hyperbolic geometry (e.g., hyperbolic neural networks [3], Poincaré embeddings [12])—into a single pipeline that reaches competitive or best-reported numbers on standard benchmarks.
    • Module roles are well separated (HSE for cross-modal alignment in hyperbolic space, GSI for geometry-aware fusion into visual features, URM for SSL stability under pseudo-labels), and ablation-style evidence in the paper supports that these pieces are complementary rather than redundant.
    • Training on a single RTX 3090 while reporting strong performance across three datasets is a notable practical strength for MICCAI-style work, where reproducibility and hardware accessibility matter for many groups.
    • The evaluation spans multiple datasets and metrics (e.g., Dice, Jaccard, boundary-oriented scores), which strengthens the claim that gains are not an artifact of a single benchmark or a single score.
  • 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 high-level idea of using text in semi-supervised medical segmentation is not new; closely related vision–language SSL methods include VCLIPSeg [7] and Text-SemiSeg [5], which already exploit textual cues. The novelty here is primarily geometric (hyperbolic alignment and fusion) plus URM, not a new problem formulation; the authors could sharpen the contrast to these lines in a single comparative paragraph.
    • The motivating example of entailment-style hierarchy (e.g., “tissue” → “organ” → “pancreas”) appears in the Introduction as intuition; the Proposed Method does not describe an explicit class taxonomy graph or ontology injected into training. Readers may therefore question how strongly the experiments support “explicit hierarchical taxonomy modeling” versus “hyperbolic space as an inductive bias for standard per-class prompts.”
    • Fairness and transparency of comparisons could be strengthened: backbones differ across dual branches (V-Net vs. ResVNet [19]), and the degree to which all baselines match preprocessing, crop sizes, and training budget should be spelled out so gains are not attributed to incidental protocol differences.
    • Computational analysis is thin relative to method complexity: dual networks, hyperbolic maps, and mixing incur overhead; wall-clock training time, memory footprint, and inference latency versus strong Euclidean VL baselines (e.g., [5,7]) would make the RTX 3090 claim easier to interpret.
    • Robustness of text conditioning is underexplored: sensitivity to prompt templates, class imbalance, modality/domain shift, and failure modes when CLIP text semantics misalign with CT/MRI appearance would strengthen clinical relevance.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    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 most impressive aspect was that most metrics at the commercial GPU level produced SoTA or near-SoTA results.

  • 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

Thanks for constructive comments. We address the main concerns below.

[@R1, R2] Curvature parameter: c is set to 0.1 in all experiments, and we will report this setting in the final version.

To R1 CLIP: HyperSemi primarily employs the frozen CLIP text encoder to provide class-level semantic guidance, rather than relying on CLIP visual features for volumetric segmentation. CoOp [32] is adopted to learn domain-adaptive continuous prompts in the form of [v_1]…[v_m][CLS_k].

URM: It substitutes uncertainty-guided for random region mixing, leveraging dual-network prediction discrepancy to locate the most uncertain 3D region where unlabeled and labeled data are mixed. It differs from other strategies.

Performance gain: The smaller gain on BraTS-2019 stems from its heterogeneous appearance, ambiguous boundaries, and strong baselines, leaving little room for improvement. More importantly, unlike the clear anatomical taxonomies in LA and Pancreas, BraTS tumor subregions lack such hierarchical structure, where hyperbolic geometry offers the most benefit.

Fairness analysis: We aim to verify the plug-in ability of HSE. Since GSI relies on hyperbolic representations derived from HSE and URM focuses on region mixing, we insert HSE only into MCF and BCP to validate whether hyperbolic semantic alignment independently benefits SSL frameworks.

To R2 Hyperbolic operation: Hyperbolic operations are intentionally applied to global pooled vectors, not 3D features. The Poincaré ball models category-level hierarchies as an inductive bias, and the aligned features are injected into volumetric representations via GSI. This indirectly refines spatial features without dense 3D operations, leveraging hyperbolic geometry precisely where it matters most for medical hierarchies.

Hyperbolic alignment : We implemented a Euclidean counterpart by replacing the hyperbolic operations in HSE and GSI with Euclidean alignment and fusion at the same feature level. On the LA dataset, the Euclidean version achieves 90.42 Dice, 82.61 IoU, 6.20 HD95, and 1.76 ASD, whereas HyperSemi reaches 91.56, 84.51, 5.13, and 1.54, respectively. On the Pancreas dataset, the Euclidean variant yields 81.48, 68.98, 6.47, and 1.38, compared to HyperSemi’s 83.68, 72.21, 5.36, and 1.25.These results consistently demonstrate the advantage of hyperbolic geometry over standard Euclidean alignment at the same feature level.

To R3 Difference: We acknowledge that text-guided semi-supervised segmentation has been explored in VCLIPSeg and Text-SemiSeg. As discussed in the Introduction, existing methods primarily perform cross-modal interaction in Euclidean space, whereas HyperSemi reformulates vision-language alignment in hyperbolic space.

Hyperbolic space: HyperSemi does not rely on an explicit class taxonomy graph or ontology. Instead, the anatomical hierarchy is encoded as an inductive bias via the geometry of the Poincaré ball, rather than being manually injected into the training process.

Fair comparison: The dual-branch design with V-Net and ResVNet follows the setting of MCF and is not introduced as an unfair advantage. As for the experimental protocol, we have reported crop sizes, optimizer, batch size, and training settings in the Implementation Details section.

Complexity: We have replaced the hyperbolic operations in HSE and GSI with Euclidean alignment and fusion at the same feature level. HyperSemi requires 76.37s/epoch, 8.31GB memory, and 2.533s/volume vs. 53.77s, 7.29GB, and 2.109s for the Euclidean version, showing moderate overhead with consistent gains.

Prompt analysis: We have added a prompt sensitivity analysis. Using the fixed prompt “A photo of [CLS]”, our method achieves 90.72 Dice, 83.12 IoU, 6.33 HD95, and 1.77 ASD. In comparison, CoOp yields 91.56, 84.51, 5.13, and 1.54, respectively. This demonstrates that learnable prompts are more robust than fixed templates when adapting CLIP text semantics to medical images.



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.

    Strong conceptual novelty in hyperbolic vision-language alignment, consistent SOTA gains across three datasets



back to top