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

Abstract

Accurate pulmonary lesion segmentation is vital for clinical diagnosis, yet conventional single-modal models often struggle to leverage the rich semantic priors available in diagnostic reports. While multimodal approaches have emerged to bridge this gap, recent research has transitioned from dense additive fusion to redundancy-reduction strategies such as token pruning. However, existing pruning-based methods remain suboptimal due to a mismatch in Token Scale—pruning concise text instead of redundant visual tokens—and a Delayed Interaction paradigm that allows background noise to propagate through the encoder. To address these limitations, we propose Text-driven Semantic Information-sieve (TeSi), an Information Bottleneck (IB) guided framework that shifts the focus to early encoder-side sieving. Specifically, to resolve the scale mismatch, an IB-Sieve mechanism treats text as a semantic filter to rank and hard-select informative visual prototypes; and to reform the interaction paradigm, our framework performs early visual sieving in the encoder, followed by an asymmetric decoder with adaptive context fusion that selectively refines textual guidance for precise boundary recovery without re-introducing noise. Extensive experiments on the QaTa-COV19-v2 and MosMedData+ datasets demonstrate that TeSi achieves competitive performance compared to current state-of-the-arts. Moreover, our method provides improved clinical interpretability through explicit prototype-to-report alignment, validating the effectiveness of the proposed sieving paradigm.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0685_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)

QaTa-COV19-v2 dataset: https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset MosMedData+ dataset: https://www.kaggle.com/datasets/maedemaftouni/covid19-ct-scan-lesion-segmentation-dataset

BibTex

@InProceedings{LonJia_Text_MICCAI2026,
        author = { Long, Jiake AND Zeng, Xinyi AND Wang, Yan},
        title = { { Text as the Sieve: Information Bottleneck Guided Prototype Sieving for Text-Guided 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 addresses two key issues in text-guided medical image segmentation: Token Scale Mismatch (where existing methods prune concise text rather than redundant visual tokens) and the Delayed Interaction Paradigm (where text guidance is introduced only at the decoding stage). To this end, the TeSi framework is proposed. Its core components include: 1.IB-Sieve Mechanism: Utilizes text semantics to rank and perform hard selection on visual prototypes, filtering out visual redundancy early at the encoder stage. 2.Asymmetric Decoder: Incorporates adaptive text fusion to prevent semantic drift. 3.Axial Prototype Generation: Compresses the sequence length from H x W to H+W . The method achieves competitive results (91.37% Dice) on the QaTa-COV19-v2 and MosMedData+ datasets.

  • 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.Accurate and Insightful Problem Formulation: The observation that existing token pruning methods target the wrong object (pruning concise text tokens instead of redundant visual ones) is valuable, with a clear motivation. The formulation of “Token Scale Mismatch” and “Interaction Paradigm” is precise. 2.Sound IB-Sieve Design: Using text as a semantic sieve to rank and hard-select visual prototypes integrates the Information Bottleneck principle with cross-modal interaction naturally. The design of aggregating residual prototypes into a background token also thoughtfully addresses information preservation. 3.Effective Axial Prototype Compression: The axial decomposition compressing H×W to H+W is a practical engineering design that balances efficiency and representational power. 4.Rationale for Asymmetric Decoder: The design of bypassing SPI to use original text embeddings directly to prevent semantic drift is logical, and ablation studies verify its superiority over symmetric designs. 5.Interpretability Analysis (Spatial Semantic Maps): The multi-scale semantic maps in Fig. 3 provide intuitive visualization of text-visual alignment, enhancing the method’s interpretability. 6.Well-Structured Ablation Studies: Table 2 covers core components such as interaction paradigms, screening mechanisms, and decoder designs, clearly showing the contribution of each module.

  • 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.Marginal Performance Gains and Lack of Statistical Significance: On QaTa-COV19-v2, TeSi (91.37%) is only 0.16% higher than FMISeg (91.21%) and 0.34% higher than ABP (91.03%). On MosMedData+, TeSi (77.96%) is notably lower than FMISeg (79.30%). Such minute differences are unconvincing without reporting standard deviations from multiple runs. The paper fails to report any statistical significance tests or variance information. 2.Narrow Dataset Coverage Limited to COVID-19 Lung Scenarios: Both datasets are COVID-19 lung images (one X-ray, one CT), making the claim of “medical image segmentation” overly broad. The lack of validation on other organs, diseases, or modalities makes it difficult to assess generalization capability. Specifically, it is unclear if the effectiveness of IB-Sieve relies on the relatively uniform background of lung images. 3.Sub-SOTA Results on MosMedData+ with Inadequate Explanation: On MosMedData+, TeSi’s Dice (77.96%) and mIoU (63.87%) are lower than FMISeg (79.30%/65.71%) and TGCAM (where mIoU is lower than TeSi). The paper attributes this to “limited training data hinders Transformer-based generalization,” yet TeSi uses a ConvNeXt backbone, not a Transformer. This explanation is self-contradictory. 4.Weak Connection to Information Bottleneck Theory: While the paper claims inspiration from the IB principle, the connection is largely intuitive. There is no formal analysis or experimental validation regarding the minimization of I(Z;X) and maximization of I(Z;Y) . IB-Sieve is essentially a similarity-based top-k selection; branding it as “Information Bottleneck” feels like over-packaging. 5.Lack of Analysis on Fixed Selection Ratio ( k=0.5 ): IB-Sieve uses a fixed 50% selection ratio, but the paper does not analyze the impact of different k values. The optimal k likely varies with lesion size and modality, making this a key hyperparameter that requires discussion. 6.Missing Computational Efficiency Comparison: The paper emphasizes reducing computational redundancy as a core motivation but fails to report efficiency metrics (FLOPs, parameter count, inference time) compared to baselines. It is impossible to judge the actual efficiency gains from axial prototype compression and IB-Sieve. 7.Negligible Contribution Differences in Ablation Studies: In Table 2, the Dice scores for Rows 2-7 range from 91.02% to 91.37%, a maximum difference of only 0.35%. This undermines the claim of “necessity” for each component—the differences may well be within the noise range. 8.Incomplete Baseline Comparisons: AT and ABP lack results on MosMedData+ (marked as “-“), weakening the completeness of cross-dataset comparisons. If this is due to unavailable code, it should be stated.

  • Please rate the clarity and organization of this paper

    Good

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

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

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

    N/A

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

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper proposes a reasonable framework for text-guided medical image segmentation. The core insight—that visual redundancy rather than text tokens should be pruned—is valuable. The design of the IB-Sieve mechanism and the asymmetric decoder is clear, and ablation studies cover major design decisions. However, the performance gains on the primary dataset are very marginal (+0.16% Dice) and lack statistical significance verification. The dataset coverage is narrow, the connection to IB theory is weak, and the results/explanation for MosMedData+ are contradictory.

  • 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 paper sits directly on the threshold. Ultimately, the balance tilts slightly toward acceptance due to the following considerations:

    • Practical Engineering Value: The rebuttal successfully demonstrates that the architectural components do exactly what they were designed to do—reduce computational complexity (25.7 GFLOPs vs. up to 54.2 GFLOPs in baselines) while maintaining highly competitive segmentation accuracy. The Axial Prototype Generator’s sequence length compression (H×W→H+W) is a useful engineering contribution.

    • Statistical Validity Established: By providing multi-run statistics (std<0.04), the authors have proven that their architectural improvements are genuine developments and not random seed noise. This satisfies a critical requirement from the Meta-Reviewer and Reviewer #1.

    • Interpretability Gains: As highlighted by multiple reviewers, the generation of intuitive, text-aligned spatial semantic maps provides strong clinical explainability that bridges the gap between black-box models and clinician adoption.

    Condition for Final Camera-Ready Upload: The authors must explicitly temper their claims in the final manuscript. They must substitute the over-packaged language of “Information Bottleneck Optimization” with “IB-inspired structural capacity constraints” as specified in their response. Furthermore, they must include the newly cited GFLOPs/parameter counts, add the standard deviations, correct the self-contradictory Transformer/CNN text, and clearly document the uniform-background lung dataset limitation in the discussion section.



Review #2

  • Please describe the contribution of the paper

    The paper identifies two main weaknesses in current text-guided medical image segmentation. First, there is a token scale mismatch, which focuses on text pruning instead of removing redundant images. Second, there is a suboptimal interaction paradigm, where late-stage fusion allows noise to spread. To tackle these issues, the authors propose TeSi, which is based on the Information Bottleneck (IB) principle for early visual redundancy suppression. The key components are the IB-Sieve mechanism, which uses text as a semantic filter to choose informative visual prototypes, and the asymmetric decoder, which prevents semantic drift by using text as a steady anchor during boundary refinement.

  • 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 work successfully uses the information bottleneck principle in medical multimodal learning to tackle visual redundancy. This is novel.

    2.The Axial Prototype Generator effectively reduces sequence length from quadratic to linear, making the computation manageable and representing a technical innovation.

    3.Unlike black-box fusion models, TeSi provides spatial semantic maps that clearly indicate which parts of the image the model focuses on based on specific text cues, such as “lower right lung.” This shows superior interpretability compared to previous works.

    4.The method used achieved a high 91.37% Dice score on the QaTa-COV19-v2 dataset and demonstrates robustness on the smaller MosMedData+ dataset, outperforming recent benchmarks like ABP and FMISeg.

  • 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 study mainly focuses on X-ray and CT for pulmonary infections. It is not clear how the “axial prototype” approach would work with 3D volumetric data or other imaging methods like MRI, where textures are more intricate.

    2.The IB-Sieve has a fixed selection ratio of k=0.5.The paper does not examine the impact of varying ratios on performance across different lesion sizes.

    3.Although the results are robust, the COVID-19 datasets are somewhat specialized. Testing on a wider range of pulmonary diseases would better confirm clinical generalizability.

  • 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?

    No

  • 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

    None

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The article deals with an important and frequently neglected problem in multimodal medical imaging: the enormous discrepancy between succinct textual descriptions and the redundant visual information. The use of preventive sieving in the early stages rather than remedial fusion in the late stages is both logical and empirically valid. The spatial semantic maps’ interpretability is especially useful to clinical adoption.

  • 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.

    Having carefully read the authors’ rebuttal, I maintain and strengthen my recommendation to accept this paper. The three concerns I originally raised, sieving ratio, limited dataset generalization, and 3D/MRI applicability, are each addressed satisfactorily: the authors report ablations across k ∈ {0.3–0.7} supporting k=0.5 as a principled choice, honestly acknowledge the COVID-19 scope limitation while clarifying that the design is not lung-specific, and appropriately scope volumetric extension as future work. The IB formalization is further clarified through an explicit variable mapping (X→Z→Y) with a credible justification for preferring structural over explicit variational optimization. The additional stability data (Dice std < 0.04 pp, ablation gaps of 0.12–0.35%) reinforces confidence in the reported improvements within the high-performance regime. The rebuttal demonstrates thorough command of the method’s theoretical grounding and empirical behavior, and the core contributions of IB-inspired early visual sieving, linear-complexity axial prototype generation, and interpretable spatial semantic maps remain compelling and well-supported.



Review #3

  • Please describe the contribution of the paper

    This paper proposes TeSi, a text-guided medical image segmentation framework that uses text as a semantic sieve to select informative visual prototypes early in the encoder. The method is motivated by two claimed limitations of prior work: Token Scale mismatch and Delayed Interaction. Concretely, the paper introduces axial prototype generation, a Semantic Prototype Interaction (SPI) module, an IB-Sieve mechanism for text-guided prototype selection, and an Asymmetric Decoder with adaptive text fusion. Experiments are conducted on QaTa-COV19-v2 and MosMedData+, with ablation studies and qualitative interpretability analysis. The overall paper is well organized and the motivation is understandable.

  • 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 identifies two intuitively meaningful issues in prior text-guided segmentation methods: pruning text rather than visual redundancy, and introducing text too late in the pipeline. This provides a coherent rationale for encoder-side semantic filtering. 2.The method is not merely a stack of unrelated modules. The prototype compression, SPI alignment, early prototype sieving, and asymmetric decoding form a consistent design narrative. 3.The paper provides component ablations and spatial semantic map visualizations.

  • 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.In section 2.3, the paper frames the method as information bottleneck guided and minimizing I(Z;X) while maximizing I(Z;Y). The manuscript does not clearly define what the X, Y and Z actually represents in this problem settings. Moreover, the implementation is text-visual similarity ranking followed by top-k selection, trained only with CE and Dice loss, without an explitic information-bottleneck objective, variational formulation or mutual-information-based optimization. The IB appears largely conceptual rather than methodologically rigorous. Can the authors clarify in what precise sense the method is an Information Bottleneck method. 2.The paper states that it reports Dice, mIoU and Accuracy, but the tables only show Dice and mIoU. Why are Accuracy results not shown? 3.In Table 2, nearly all multimodal variants outperform the uni-modal baseline by roughly 4 points in both Dice and mIoU, while the differences among the proposed components themselves are relatively small. This makes it difficult to determine the specific effect size of each design choice. The ablation mainly shows that introducing the overall multimodal framework is beneficial, but does not convincingly disentangle how much of the gain comes from IB-Sieve, SPI, the spatial map recovery, or the asymmetric decoder individually. Can the authors clarify what conclusions can be drawn from Table 2? 4.In section 2.1 and 2.2, the paper claims reducing redundancy and improving efficiency. however, these efficiency benefits are not quantitatively validated in the experiments.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper addresses an interesting problem and proposes a coherent multimodal segmentation framework. However, I recommend rejection because the main claims are not sufficiently supported by rigorous evidence. The experimental comparison does not appear to be based on a clearly unified reproduction of baselines, which weakens the claimed performance advantage. Moreover, the Information Bottleneck interpretation is insufficiently defined, and the ablation study does not clearly isolate the effect of each proposed component. As a result, the paper’s central contributions remain only partially validated.

  • 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.

    After reading the authors’ rebuttal, I appreciate their detailed responses, which have addressed most of my concerns. I am therefore willing to raise my score to WR. However, I remain unconvinced by the explanation of the IB component, particularly the interpretation of minimizing I(Z;X). Overall, the rebuttal improves my confidence in the work, but this remaining conceptual concern prevents me from assigning a more positive score.



Author Feedback

We sincerely thank the reviewers for their constructive comments. Our responses are below: Q1 IB formalization (R1, R3) We appreciate this critical comment. Our method follows an IB-inspired structural bottleneck design rather than explicit variational IB optimization. Using our paper’s notation, the IB variables correspond as: X: multi-scale features {Xi} Z: sieved high-semantic prototypes Z_sem Y: segmentation mask (E_text as a semantic proxy for Y) A structural IB approximation is induced via capacity constraints on Z (see Q4), jointly promoting: Min I(Z;X): Axial compression reduces token capacity (↓H(Z)); IB-Sieve further prunes P_v to Z_sem, discarding task-irrelevant regions into the p_bg token. Max I(Z;Y): SPI aligns prototypes across modalities, ensuring Z remains predictive of Y, further reinforced by Dice/CE objectives. Our early trials with explicit IB loss led to instability with marginal gains, suggesting the structural design is sufficient in practice. Q2 Model effectiveness and module contributions (R1, R3) We respectfully clarify that the modest improvements reflect a high-performance region (>91% Dice), where only incremental gains are expected. Similar patterns also appear in prior works[16,23], including their ablations. We agree that reporting standard deviations would strengthen the results. Across 8 runs, our TeSi shows stable performance with Dice std < 0.04 pp on QaTa, while ablation gaps (0.12%–0.35%) exceed this variation range, suggesting reliable improvements. Table 2 isolates each module via targeted substitutions to disentangle specific contributions: IB-Sieve: removing sieving (Row 4) leads to the lowest performance, indicating its importance. SPI: replacing with MHCA (Row 5) results in a consistent drop, supporting the shared alignment design. Spatial map: Row 6 shows a similar decreasing trend. Decoder: Rows 2/3/7 support the asymmetric fusion design; a decoder-only variant reduces to a late-fusion scheme akin to ABP[16]. Q3 Missing Accuracy and incomplete baselines (R1, R3) We apologize for the confusion. The mention of ACC was a typo. ACC is near ceiling on this task (e.g., >97% on QaTa) and omitted in recent works[11,16,23], making it less informative for comparison. We will correct this. All baseline results are taken from original papers under their reported settings with identical dataset splits, following standard practice. ABP and AT lack MosMedData+ results due to unavailable code or missing task configurations; we avoid potentially unfair re-implementation and will clarify this in the final version. Q4 Computational efficiency (R1, R3) By compressing encoder tokens from 4165 to 147 (~28×), our design drastically cuts cross-modal encoder cost, achieving improved results with only 25.7 GFLOPs (vs. 54.2/17.7/34.3 for [9,10,23]) and ~75% of FMISeg[23]’s parameters (text encoder excluded). Q5 Sieving Ratio (R1, R2) We explored k ∈ {0.3, 0.4, 0.5, 0.6, 0.7}. Early checks suggested no clear advantage of lesion/dataset-specific tuning, so k=0.5 was chosen as a fixed midpoint to avoid overfitting. Adaptive ratios are interesting for future work. Q6 MosMedData+ results (R1) We clarify that TeSi slightly outperforms TGCAM but remains below FMISeg on MosMedData+, and our claim is therefore “competitive”, not SOTA. The confusing wording refers to the attention-based SPI module rather than the CNN backbone, and will be revised. Following prior work[10], we used mild augmentation, while FMISeg adopts heavier strategies (e.g., rotation, noise) according to its code, which may partly explain the performance gap. Q7 Dataset scope and Generalization (R1, R2) We acknowledge the current limitation to COVID-19 lung datasets, due to limited availability of public datasets with paired reports. The core design—text-driven visual sieving—does not assume lung-specific properties. The strategy can in principle be extended to 2.5D/3D prototypes, but validation on MRI and volumetric data remains 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.

    This paper studies text-guided medical image segmentation and proposes the TeSi framework to address two limitations of existing methods: the scale mismatch between concise textual tokens and redundant visual tokens, and the delayed interaction caused by introducing textual information only at the decoding stage. Overall, the topic is relevant and meaningful, the motivation is reasonably clear, and the framework is coherently structured.

    However, the reviews show considerable divergence, as several central claims of the paper still require stronger experimental evidence and more rigorous theoretical justification. In the rebuttal, the authors should clearly explain the precise relationship between IB-Sieve and the Information Bottleneck principle; provide or justify the absence of standard deviations over multiple runs and statistical significance analysis; clarify the actual contribution of each proposed module; report efficiency-related metrics to verify whether axial prototype compression and visual prototype sieving indeed bring computational benefits; discuss the effect of different sieving ratios across lesion sizes and datasets; clarify whether baseline comparisons were conducted under a unified experimental setting, and explain the missing Accuracy results and incomplete baseline results; and discuss the potential applicability and limitations of the method beyond COVID-19 lung datasets, including other organs, MRI, and 3D volumetric data.

  • 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 addresses a meaningful problem in text-guided medical image segmentation by focusing on visual-token redundancy and the delayed use of textual guidance in existing methods. The proposed TeSi framework is coherent, with axial prototype generation, text-guided visual sieving, and an asymmetric decoder forming a clear design around early semantic filtering. The initial concerns about marginal gains, weak connection to the Information Bottleneck principle, missing efficiency evidence, fixed sieving ratio, and limited dataset scope were valid. However, the rebuttal addresses these issues in a reasonably convincing way. In particular, the authors clarify that the method is IB-inspired rather than a formal variational IB optimization, provide multi-run stability results, report computational cost showing meaningful efficiency advantages, and add analysis of the sieving ratio. Although the evaluation remains limited to COVID-19 lung datasets and the MosMedData+ results are not uniformly superior to all baselines, the method is technically sound, interpretable, and practically useful. Overall, I believe the paper makes a solid contribution to text-guided medical segmentation and would support acceptance, with the expectation that the final version clearly tempers the IB claim and discusses the dataset-scope limitation.



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.

    The paper proposes TeSi, a text-guided segmentation framework that performs early visual prototype sieving using textual semantics, combined with axial prototype compression and an asymmetric decoder. Reviewers agreed that the motivation is clear and that the method forms a coherent design addressing visual redundancy and late text interaction. Initial concerns included marginal gains, lack of variance/efficiency reporting, weak formal connection to Information Bottleneck theory, fixed sieving ratio, and limited validation on COVID lung datasets. Limitations remain: the MosMedData+ result is competitive but not best, and generalization beyond pulmonary datasets is unproven. Nevertheless, the method is technically sound, interpretable, and sufficiently validated for acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The paper presents a useful framework for text-guided medical image segmentation, with meaningful early visual sieving and interpretable spatial semantic maps. However, the newly added results cannot be considered during rebuttal per MICCAI policy.



back to top