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

Abstract

Vision-Language Model (VLM) has great potential to enhance the quality of pseudo labels in semi-supervised spine segmentation by leveraging textual class prompts to generate segmentation map, but no one has studied it yet. Although promising, it lacks explicit constraints to ensure consistency between spine class prompts and spine unit region, resulting in unsatisfactory performance in multi-class segmentation map generation. In this paper, we propose CPS4, the first text-guided semi-supervised spine segmentation network using class prompts to enhance the quality of spine pseudo labels. Specifically, CPS4 is implemented through two training stages. (i) Class-specific consistency constrained VLM pretraining stage: we propose token- and pixel-level attention loss to optimize the consistency between class prompts and spine units, forcing the textual class prompt to be closely coupled with the target spine unit in the semantic space. (ii) Class Prompt driven semi-supervised spine segmentation stage: using the pretrained vision–text encoder, we derive each class-specific binary segmentation map for the unlabeled spine image and integrate them into an unified multi-class segmentation map, improving the quality of the spine pseudo label generated by the semi-supervised spine segmentation network. Experimental results show that our CPS4 achieves superior spine segmentation performance with Dice of 80.44%, only using 5% labeled data on the public spine segmentation dataset, surpassing popular semi-supervised learning and VLM methods. Our code will be available.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0946_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/QingtaoPan/CPS4

Link to the Dataset(s)

https://www.spinesegmentation-challenge.com

BibTex

@InProceedings{PanQin_CPS4_MICCAI2026,
        author = { Pan, Qingtao AND Sun, Hongzan AND Ji, Bing AND Li, Shuo},
        title = { { CPS4: Class Prompt Driven Semi-supervised Spine Segmentation with Class-Specific Consistency Constraint } },
        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 manuscript proposes CPS4, a novel framework for semi-supervised spine segmentation that leverages Vision-Language Models (VLMs) to enhance pseudo-label quality. Addressing the limitation of existing VLMs which lack explicit constraints between textual class prompts and specific anatomical regions, the authors introduce a two-stage training strategy.

    1. Stage 1 (Pretraining): A class-specific consistency constrained VLM pretraining stage is introduced, utilizing proposed token-level and pixel-level attention losses to enforce tight semantic coupling between class prompts (e. g. , “Lumbar Vertebra 1”) and their corresponding spine units in the feature space.
    2. Stage 2 (Segmentation): In the semi-supervised segmentation stage, the pretrained vision-text encoder generates class-specific binary segmentation maps for unlabeled images. These maps are aggregated into a unified multi-class pseudo-label to guide the training of a Mean Teacher-based method.
  • 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 integration of textual class prompts into semi-supervised spine segmentation is a significant and timely contribution. While VLMs have been explored in medical imaging, this work is the first to explicitly address the “cross-class interference” problem in multi-class spine segmentation by designing class-specific consistency constraints (token- and pixel-level attention losses). This moves beyond simple image-text alignment to fine-grained region-prompt binding. 2.The methodological design is logical and well-grounded. The two-stage approach effectively decouples the challenging task of learning robust vision-language alignment from the segmentation optimization. 3.Spine segmentation is critical for clinical applications but suffers from high annotation costs due to the multi-class nature (vertebrae + discs). By demonstrating superior performance with only 5% labeled data (achieving ~80.44% Dice), this work offers a practical solution to reduce the reliance on expensive expert annotations.

  • 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 method relies heavily on the quality and specificity of the textual class prompts (e. g. , “A magnetic resonance image of a {class name}”). The paper does not discuss the sensitivity of the model to prompt engineering variations (e. g. , synonyms, different sentence structures). In clinical settings, terminology can vary; the robustness of the method to such linguistic shifts is not evaluated. 2.The proposed framework requires a dedicated pretraining stage (Stage 1) before the actual semi-supervised segmentation training (Stage 2). The paper lacks a detailed analysis of the total computational cost (GPU hours) compared to end-to-end SSL methods. For practical deployment, the efficiency of generating 20 separate attention maps (for 20 classes) per unlabeled image during Stage 2 inference/training needs to be quantified. 3.The experiments are conducted solely on the MRSpineSeg dataset. While the results are impressive, the method’s generalizability to other modalities (e. g. , CT spine scans) or other multi-class anatomical structures (e. g. , abdominal organs) remains unproven. A cross-modality experiment or a discussion on potential domain shifts would strengthen the claims. 4.The dataset description notes that “Not all images contain 19 spinal structures. “ The method generates attention maps for all 20 classes for every unlabeled image. The paper does not explicitly explain how the framework handles cases where a specific vertebra is absent in the field of view. Does the VLM generate false positive attention maps for missing classes, and if so, how does the aggregation step mitigate this noise in the pseudo-labels?

  • 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 paper is well-organized and clearly written. The motivation for addressing cross-class interference in VLM-based pseudo-labeling is compelling and addresses a genuine gap in the literature. The experimental results are robust, showing consistent improvements across different labeling ratios. While the computational cost and modality generalization are minor concerns, the core contribution of class-specific consistency constraints for multi-class SSL is significant and well-executed

  • 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 author addressed the concerns



Review #2

  • Please describe the contribution of the paper

    This paper proposes CPS4, a novel semi-supervised spine segmentation framework that leverages textual class prompts to generate high-quality pseudo labels. Methodologically, it introduces a class-specific consistency constraint using token- and pixel-level attention losses during VLM pre-training to ensure strict semantic alignment and prevent cross-class interference. Extensive experiments demonstrate that CPS4 achieves state-of-the-art performance on spine MRI datasets across various labeled ratios.

  • 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.

    Overall, the motivation for this work is clear, the authors have well articulated their methodology, and the proposed approach has achieved promising experimental 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.

    The method heavily relies on class prompts, and the paper explicitly treats prompt-object consistency as the central issue. However, the prompt design is hard-coded to a single template, namely “A magnetic resonance image of a {class name},” without any prompt sensitivity analysis. As a result, it remains unclear whether the reported gains are robust to different textual formulations or are partially tied to this specific handcrafted prompt.

  • 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

    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 refer to the strengths and weaknesses above.

  • 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

    The authors address the challenge of label scarcity in multi-class spine segmentation (Vertebral Bodies and Intervertebral Discs) by utilizing textual class prompts to guide a semi-supervised learning (SSL) process. The CPS4 framework operates in two stages. In the first stage, a vision-text encoder is trained using proposed token-level and pixel-level attention losses to ensure tight semantic coupling between specific spine class prompts and their corresponding image regions. In the second stage, the pretrained VLM generates class-specific binary maps for unlabeled data, which are aggregated into unified multi-class pseudo-labels to supervise a student-teacher segmentation network.

  • 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.This appears to be the first work to explicitly introduce textual class prompts into the semi-supervised spine segmentation domain.

    2.The introduction of specific consistency constraints, namely the token-level and pixel-level losses, directly addresses the known issue of VLMs failing to focus on small or specific anatomical objects. The token-level loss aggregates attention activations toward the predicted spine unit, and the pixel-level loss prevents over-aggregation into subregions, ensuring the VLM covers the entire target unit.

    3.The model achieves a Dice score of 80.44% using only 5% labeled data. It consistently outperforms established baselines like Mean Teacher (MT), MedCLIP, and MGCA across various labeling ratios (5%, 10%, 25%, 50%).

    4.The ablation study clearly demonstrates that both token-level and pixel-level losses provide complementary gains, with the full model significantly outperforming the baseline.

  • 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 authors acknowledge that the MRSpineSeg dataset consists of volumetric (3D) images. However, the proposed architecture uses a 2D ResUNet. Spine anatomy is inherently 3D, and spatial consistency between adjacent slices is critical for identifying specific vertebrae (e.g., distinguishing T12 from L1). By treating slices as independent 2D inputs, the model ignores the rich volumetric context provided by the MRI. This is a step backward from the current state-of-the-art medical image segmentation, which typically employs 3D U-Net variants or Transformers to maintain anatomical continuity. The authors should justify why a 3D baseline (like nnU-Net) or a 2.5D contextual approach was not employed, and provide a comparison to prove that their text-guided pseudo-labels compensate for the lack of 3D spatial information.

    2.The method generates pseudo-labels by “stacking” binary maps from the VLM and adding them to teacher predictions. However, the authors do not explain how they handle label conflicts. What happens if a pixel is predicted as both “L1” and “L1/L2 disc” by different class-specific maps? A simple summation of two potentially misaligned probability maps (in the loss function) does not seem mathematically sound - if both models are confident, the resulting “label” could exceed valid probability ranges or create highly noisy gradients. A more robust approach, such as Bayesian fusion or confidence-weighted averaging, is missing.

    3.The method for aggregating 19+ class-specific attention maps into a single pseudo-label lacks a robust mechanism for handling overlapping predictions or false positives in slices where certain structures are absent. If the VLM is forced to produce a “T10” segmentation map for a slice that only contains “L5,” it will likely pick up noisy, similar-looking features. Stacking these “hallucinated” structures into the final pseudo-label could introduce significant systematic errors that the student network will then memorize.

    4.The method relies on OTSU automatic thresholding to binarize attention maps into pseudo-labels. However, the paper lacks a sensitivity analysis or discussion on whether OTSU remains robust across different MRI qualities or if it tends to “over-segment” in noisy images.

    5.Statement like “but no one has studied it yet.” (abstract). Such absolute statements can be risky; it is often safer and more professional for authors to claim to be “among the first” or “to the best of our knowledge.”

  • 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?

    CPS4 offers a novel VLM-based approach to semi-supervised spine segmentation with impressive low-label performance (80.44% Dice at 5% labels), yet it faces significant technical hurdles. The use of a 2D architecture for volumetric data ignores essential 3D spatial context and lacks comparison with standard 3D baselines like nnU-Net. Additionally, the pseudo-label aggregation process is mathematically vague, providing no clear strategy for resolving label conflicts or preventing hallucinations in slices where specific anatomical structures are missing. Without a robustness analysis for its thresholding methods and a justification for its 2D/3D trade-offs, the paper’s current methodological rigor does not fully meet the standard for acceptance at this stage.

  • 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 successfully addressed my previous concerns regarding the model architecture and experimental validation during the rebuttal. The proposed method is sound and outperforms the baseline models. I am happy to recommend acceptance.



Author Feedback

Thanks to all reviewers for your careful reviews and appreciation of our work, such as novel framework (R1R2R3), significant contribution (R1), compelling motivation (R1R2), superior experiment (R1R2R3), and well writing (R1R2).

Q1: Prompt Robustness (R1R2). A1: CPS4 is robust to synonyms and different sentence structures. For sentence structures, the prompts “A magnetic resonance image of a {class name}”, “A spine MR image containing {class name}”, and “The anatomical structure of {class name} in MRI” achieved mDice of 80.44, 80.35, 80.47.For synonyms, removing “Vertebra” or replacing it with “Spine”, e.g., “Lumbar Vertebra 1”, “Lumbar 1”, and “Lumbar Spine 1”, achieved mDice of 80.44, 80.32, 80.39.These small fluctuations demonstrate the robustness of CPS4. Q2: Efficiency Analysis (R1). A2: CPS4 requires no extra inference cost, since attention maps are generated only during training. On an NVIDIA Tesla V100 GPU, both CPS4 and baseline take 18 ms/image during inference. For training, CPS4 takes 13h34min in total, while the baseline takes 9h31min. The additional cost includes 2h31min for stage 1, 1h19min for attention map generation in stage 2, and 13min for the loss backpropagation.​ Despite additional cost, CPS4 brings a 6.19 mDice improvement under 5% labeled data.

Q3: Generalizability (R1R3). A3: This work focuses on 2D spine segmentation. In our another work, we extended CPS4 and evaluated it on 3D semi-supervised spine segmentation, including 3D organ and 3D spine tasks. Using 3D nnUNet as the backbone, under 10% labeled data, CPS4 achieves 82.63 mDice on MACT organ dataset, outperforming MT and CLIP by 10.93 and 3.69; on MRSpineSeg dataset, CPS4 achieves 73.44 mDice, outperforming MT and CLIP by 7.07 and 3.81.Using 3D ResUNet as the backbone, under 10% labeled data, CPS4 achieves 82.06 mDice on MACT organ dataset, outperforming MT and CLIP by 8.95 and 2.33; on MRSpineSeg dataset, CPS4 achieves 72.64 mDice, outperforming MT and CLIP by 6.59 and 4.25.These results validate CPS4’s generalizability across different segmentation scenarios.

Q4: Absent-Class Handling (R1R3). A4: After OTSU divides each attention map into foreground and background where each attention map is normalized to [0,1], we compute mean attention score of the foreground region. If this mean score is lower than 0.5, the class is regarded as absent, and its binary map is set to background. This reduces false-positive activations from missing classes. We will clarify this in the revised manuscript.

Q5: Label Conflicts and Probability Maps Summation (R3). A5: For label conflicts, if a pixel is predicted as different classes after OTSU, we select the class with the highest score in attention maps, producing a non-overlapping text-guided label. For example, if both L1 and L1/L2 are activated at one pixel, but their original attention scores are 0.82 and 0.65, this pixel is assigned to L1.For probability maps summation, we add teacher prediction and text-guided label, normalize the added score, and apply argmax to obtain a one-hot pseudo label. For example, if teacher prediction is [0.7,0.2,0.08,0.02] and text-guided label is [1,0,0,0], their summation is [1.7,0.2,0.08,0.02]. After normalization, it becomes [0.85,0.1,0.04,0.01]. The final one-hot label is [1,0,0,0]. We will add these explanations in the revised manuscript and further evaluate Bayesian fusion and confidence-weighted averaging to assess their potential benefits over our fusion strategy.

Q6: Sensitivity Analysis of OTSU (R3). A6: OTSU is robust to different MRI qualities. Gaussian noise is added to unlabeled MRI inputs during the text-guided label generation process in Stage 2.Standard deviations were set to 0.05, 0.10, 0.15, and CPS4 achieves mDice scores of 80.44, 80.31, 80.12, under 5% labeled data. Compared with the original result of 80.44 without noise, performance drops are only 0.00, 0.13, 0.32.These minor decreases indicate that OTSU is robust to different MRI qualities.



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.

    All reviewers are inclined to finalize their decisions following the rebuttal. The authors are advised to address the concerns raised by Reviewers 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.

    All reviewers have reached a consensus in favor of accepting the paper.



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.

    R1 changed their recommendation from Weak Accept to Accept, while R3 changed from Weak Reject to Accept. Both reviewers were satisfied that the rebuttal addressed their concerns with the initial version of the paper.



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.

    After the rebuttal, all reviewers agreed that the paper should be accepted. The Area Chair supports this consensus and recommends acceptance.



back to top