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

Abstract

Early detection of Congenital Heart Disease (CHD) relies heavily on subjective ultrasound interpretation. While AI offers automation, privacy regulations and data fragmentation remain major barriers. Federated Learning (FL) enables privacy-preserving collaboration, yet deploying pre-trained vision-language models like CLIP in non-IID settings is challenged by inter-hospital heterogeneity and semantic overfitting to site-specific biases. To address these issues, we propose FedCHDP, a novel federated concept-guided dual-modal prompting framework. On the textual side, we construct a federated medical concept library that decomposes diagnostic queries into fine-grained clinical descriptors (i.e., chamber numbers, flow numbers and flow patterns), enforcing semantic consensus across clients. An uncertainty-aware gating mechanism further adaptively fuses dynamic concept cues with global priors to rectify ambiguous predictions. On the visual side, we introduce a global-guided contrastive alignment that calibrates visual prompts using disease prototypes, aligning local embeddings with global semantic centers to mitigate distribution shifts. Experiments show that FedCHDP significantly outperforms state-of-the-art methods for multi-center CHD identification.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1712_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{GuRui_FedCHDP_MICCAI2026,
        author = { Gu, Ruilin AND Li, He AND Huang, Wenke AND Tan, Qingxiong AND Ye, Mang AND Du, Bo},
        title = { { FedCHDP: Federated Concept-Guided Dual-Modal Prompting for CHD Diagnosis } },
        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 FedCHDP, a federated vision–language framework for first-trimester CHD diagnosis from fetal cardiac ultrasound. On text side, it introduces a federated medical concept library that decomposes class prompts into fine-grained descriptors and fuses a concept-conditioned prompt with an unconditional disease prompt via an uncertainty-aware gate. On the visual side, it aggregates visual prototypes into global disease prototypes and uses a contrastive alignment to mitigate cross-site distribution shifts. Experiments on a 4-center dataset show gains over recent 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 concept-guided textual prompting, which explicitly encodes clinically interpretable dimensions, is well motivated for CHD subtyping. 2.The uncertainty-aware gating between unconditional disease priors and instance-conditioned concepts is a simple yet practical mechanism that is well aligned with clinical reasoning. 3.The figure is clear and lovely.

  • 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 writing needs further refinement, particularly in the method section, where the notation is difficult to follow. For example, in Section 2.1, $C$ denotes the class and $c_i$ denotes a concept, but the meaning of $K_m$ is unclear. It is also unclear what $c_{i,m}$ represents and whether it is a one-hot indicator. 2.The figures do not clearly annotate the variables, which further limits readability. 3.The prototype aggregation strategy is relatively naive, relying on sample-size-weighted averaging, which may introduce prototype bias under severe class imbalance or in the presence of low-quality clients. 4.The dependence on manually annotated concept labels may limit the applicability and scalability of the method.

  • 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

    Q1.How are the semantic anchors instantiated for each concept value? Are they defined using fixed templates or curated from clinical guidelines?

    Q2.What is the visual prompt in Section 2.3? Its definition is currently unclear.

    Q3.Could the authors report leave-one-center-out results to better assess cross-site generalization? If this is not feasible, please at least provide per-center test performance and cross-site confusion matrices.

    Q4.Would ultrasound-specific pretraining for the vision encoder change the comparative landscape versus CLIP-based approaches?

  • 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?

    My recommendation is mainly based on concerns about clarity and methodological strength. The writing is not sufficiently clear, and several key concepts and notations are inadequately defined, which makes the method difficult to follow. In addition, the prototype aggregation strategy is relatively weak, as it relies on a simple averaging scheme that may be vulnerable to bias under class imbalance or low-quality client updates.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The authors have addressed most of my concerns. Their rebuttal and the newly reported results help clarify the model design and support its effectiveness.

    If the paper is accepted, I would strongly encourage the authors to further improve the presentation and organization of the manuscript, as the current writing appears somewhat rushed and is occasionally difficult to follow. In addition, since the study relies on private data, making the dataset publicly available in the future, if possible, would be highly valuable to the community and would also enhance reproducibility.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a novel federated dual-modal prompt tuning framework based on CLIP for congenital heart disease prediction. To ensure semantic consistency across clients, the authors construct a federated medical concept library utilizing fine-grained clinical descriptors. For the textual prompt, an uncertainty-aware gating mechanism is introduced to guarantee prediction stability in ambiguous cases. Meanwhile, for the visual prompt, the method aligns local class prototypes with global ones via contrastive learning to effectively mitigate data distribution shifts.

  • 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 proposed method systematically addresses the challenges of using CLIP for clinical CHD prediction in a federated learning setting from both textual and visual modalities. Specifically, it identifies the semantic gap in the text modality and the distribution gap in the image modality, and provides tailored strategies for each. For the textual modality, it constructs a global concept library with fine-grained clinical descriptors; for the visual modality, it introduces a contrastive learning mechanism to mitigate distribution shifts. 2.The proposed uncertainty-aware gating mechanism effectively fuses the prediction of conditional predicted concepts with unconditional generic knowledge to mitigate prediction uncertainty in ambiguous cases. This mechanism is not only methodologically novel but also holds substantial clinical significance. 3.The empirical evaluation is comprehensive, showing significant improvements over baselines. The effectiveness of the proposed textual and visual components is well-supported by ablation studies. Additionally, the inclusion of t-SNE visualizations (demonstrating category-wise clustering and distinct decision boundaries), along with hyperparameter sensitivity and model convergence analyses, makes the experimental claims highly convincing.

  • 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 overall description of the method’s pipeline is not entirely clear, and several details, particularly in Figure 2, require further optimization. For instance, the figure does not explicitly denote which components (e.g., the authors can use a sign on the visual prompt) are trainable during the process. Furthermore, for classification loss, $L_{cls}$ , it remains ambiguous whether this loss measures the similarity between the text embedding and the image prototype, or the text embedding and the image embedding. Although the figure appears to illustrate similarity with the prototype, using the image embedding seems more reasonable. Finally, the methodology section fails to formally formulate the model aggregation process, especially regarding the visual prompt and the textual concept library. The variables $V$ and $T$, denoting the visual and textual encoders respectively, are defined in the text but are never formally utilized in the subsequent equations or methodological formulations.
    2. The proposed method is a dual-modal prompt tuning framework. However, many of the compared baselines—such as CoCoOp, KgCoCoOp, FedTPG, PromptFL, and FedVPT—are strictly single-modal (either textual or visual). This asymmetric comparison raises concerns regarding experimental fairness. It remains ambiguous whether the reported performance gains stem from the proposed methodological innovations or simply from the additional information provided by the second modality.
    3. Addressing non-independent and identically distributed (non-IID) data is a fundamental challenge in federated learning, particularly for classification tasks. However, the authors fail to evaluate their method under severe label skew scenarios. Specifically, they do not employ standard techniques, such as the Dirichlet distribution, to simulate non-IID label distributions across different clients. This significant omission leaves the robustness and generalizability of the proposed framework in highly class-heterogeneous environments highly questionable.
    4. The proposed method involves multiple types of information exchange between clients and the server, such as visual prompts, a medical concept library, and prototypes. In addition, the server is required to compute global prototypes, and each client needs to optimize multiple loss functions. However, the paper does not provide a thorough analysis or comparison of the resulting communication and computational overhead. This lack of evaluation may limit the practical applicability of the method in real-world clinical scenarios.
    5. The experimental results lack the reporting of variance or standard deviations across multiple runs. This omission makes it difficult to assess the true robustness and stability of the proposed model, as the reported performance improvements could be attributed to random seed variations.
    6. Please correct minor typographical errors throughout the paper to improve its rigor. Examples include “fused text enbedding” in Figure 2 and “boundaries..” in the Figure 3 caption.
  • Please rate the clarity and organization of this paper

    Satisfactory

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

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

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

    N/A

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

    N/A

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

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

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

    The paper presents a meaningful and fairly well-validated federated dual-modal prompt tuning framework for clinical CHD prediction. The method is motivated and empirically strong, and the uncertainty-aware fusion mechanism is particularly interesting. However, concerns remain regarding methodological clarity, fairness of baseline comparisons, robustness under stronger non-IID settings, and the lack of communication/computation overhead analysis. Overall, I view the paper as slightly above the acceptance threshold for MICCAI.

  • 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 rebuttal, I believe the authors have addressed most of the major concerns raised by the reviewers. In particular, they clarified the notation and methodological pipeline, explained the roles of semantic anchors and visual prompts, and provided additional experimental evidence, including variance over multiple seeds, stronger baseline comparisons, leave-one-site-out evaluation, and communication/computation cost analysis. These responses substantially improve the clarity, completeness, and credibility of the work.

    Some concerns still remain. The technical novelty is somewhat incremental, as the method mainly integrates concept-guided prompting, uncertainty-aware fusion, and prototype-based alignment. The reliance on fine-grained clinician-annotated concepts may also limit scalability, and the sample-size-weighted prototype aggregation may still be suboptimal under extreme client imbalance. Nevertheless, given the clinically meaningful problem setting, the coherent framework design, the strong empirical results, and the fact that the rebuttal has resolved most reviewers’ concerns, I believe the paper is marginally above the acceptance threshold.



Review #3

  • Please describe the contribution of the paper

    This paper proposes FedCHDP, a federated dual-modal prompting framework for CHD diagnosis based on vision-language models. The key idea is to address semantic ambiguity and cross-site distribution shift in federated settings via: 1.A concept-guided textual prompting mechanism, which decomposes disease semantics into fine-grained clinical concepts. 2.An uncertainty-aware gating module to fuse global and instance-specific predictions. 3.A global-guided contrastive alignment strategy to calibrate visual prompts using aggregated disease prototypes. The method is evaluated on a multi-center fetal echocardiography dataset and shows consistent improvements over several prompt-learning and federated 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.This paper has clear motivation and a well-defined problem. The paper clearly identifies two key challenges in federated VLM: semantic ambiguity in prompt learning and distribution shift across clients. The problem setting is clinically meaningful. 2.This paper has technically well-structured framework. The dual-modal design (text + vision) is logically consistent. Concept-guided prompting is intuitive and clinically interpretable. Also, the uncertainty-aware gating is a nice touch and mimics clinical reasoning. 3.This paper has strong experimental results. It has significant improvement over baselines. It also has done comprehensive experiments include multi-class evaluation, ablation study, t-SNE visualization and hyperparameter analysis.

  • 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.It limits novelty in core components. Many components are incremental combinations of existing ideas such as prompt learning, prototype alignment and uncertainty gating which has been widely used. Also, the novelty mainly lies in integration rather than fundamentally new methodology. 2.There are several weaknesses in the assumption of concept annotation. The method relies on fine-grained concept labels, however; in practice such annotations are expensive and not always available. Also, the paper does not discuss the robustness to noisy annotations. 3.The experimental comparison raises some fairness issues. Specifically, several baselines (e.g., ZS-CLIP) are relatively weak and may not provide a sufficiently strong reference point. In addition, the paper lacks comparisons with more recent federated vision-language learning methods, as well as non-prompt-based federated baselines (e.g., FedAvg with a strong backbone). More importantly, it remains unclear whether the reported performance gains primarily stem from the proposed methodological design or from the additional supervision introduced by fine-grained concept 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 submission does not provide sufficient information for reproducibility.

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

    N/A

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

    N/A

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

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

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

    The paper addresses an important and clinically relevant problem and presents a well-designed framework that integrates concept-guided prompting and prototype-based alignment in a federated setting. The experimental results are strong and demonstrate consistent improvements over existing methods. However, the overall novelty is somewhat limited, as the method mainly combines existing techniques rather than introducing fundamentally new ideas. Additionally, the reliance on fine-grained concept annotations and limited discussion of reproducibility weakens the impact. Overall, the paper is solid and practically meaningful, but could be strengthened in terms of novelty justification and experimental completeness.

  • 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

We thank the reviewers for the insightful feedback. We appreciate Reviewer#2(R2) and Reviewer#3(R3)’s recognition and greatly value Reviewer#1(R1)’s constructive critiques. We hope our sincere response and dedicated analyses resolve these concerns.

(1) Notation and Pipeline(R1,R2). We refined Sec. 2 and corrected typos. Specifically, $K_m$ is the number of discrete concept values for concept $m$, where $m\in{ch,fl,pa}$ denotes chamber number, flow number, and flow pattern. For example, flow number has two values (1 or 2), so $K_{fl}=2$. $c_{i,m}\in{1,\ldots,K_m}$ denotes the ground-truth discrete concept label of sample $i$ for concept $m$, not a one-hot vector. Semantic anchors are fixed embeddings from standardized templates (e.g.,``single blood flow’’) via the frozen text encoder. They provide stable references for concept prediction, while concept library embeddings are learnable.

The visual prompt $P_v$, a set of learnable parameters, is injected into the frozen visual encoder $V$ alongside the input image $x_i$, producing the local visual embedding $v_i=V(x_i,P_v)$. Textual features are extracted by the frozen text encoder $T$. Classification is performed by image-text matching: $v_i$ is matched with unconditional and concept-conditioned text embeddings to obtain $l_{up}$ and $l_{ccp}$, which are fused into $l_{fused}$ by the uncertainty-aware gate, for ${L}{cls}=CE(l{fused},y_i)$. Global prototypes are solely for contrastive alignment, not classification. Fig.2 is updated to explicitly mark frozen/trainable modules and separate these paths.

(2) Prototype Aggregation(R1). We acknowledge sample-size weighting may be suboptimal under extreme imbalance clients. Our aggregation is class-wise rather than mixing all classes. The motivation is that prototypes from small local class subsets can have high variance; weighting by sample number reduces the influence of unstable estimates. Moreover, global prototypes are not used for classification but for visual prompt regularization. Since $L_{align}$ is controlled by $\lambda_{align}$, potential prototype bias cannot dominate the primary classification objective. We will discuss class-balanced/reliability-aware extensions in the revision.

(3) Concept annotations(R1,R3). Using clinician-annotated concepts based on medical guidelines strictly aligns our model with real-world diagnosis. While incurring training costs, annotations are not required for inference. FedCHDP takes only images as input, automatically predicts clinical concepts, and constructs the conditioned prompt via a reusable concept library. Moreover, our uncertainty-aware gate ensures robustness against noisy predictions: if concept confidence is low, it dynamically down-weights the concept-guided branch.

(4)Experimental Details(R1,R2,R3). We added missing experimental details: Label Skew(R2): Our dataset inherently exhibits severe natural heterogeneity(838–1532 samples/center). For instance, the HV class has 2 cases in one center but 393 in another, and FSV ranges from 77 to 682 cases, confirming a strictly non-IID setting.

Variance&Baseline Fairness(R2,R3): Across 3 random seeds, FedCHDP achieves a stable 64.42±3.66 macro-F1.For fairness, we evaluated a non-prompt FedAvg(59.26±4.45)[AISTATS’17] and recent dual-modal FedMVP(31.14±3.90)[ICCV’25]. FedCHDP significantly outperforms them and other strong baselines like IVLP(53.02±4.42) and MaPLe(55.41±4.91). This confirms our gains arise from dynamic fusion, not just dual modalities.

Cross-Site Generalization(R1): In a leave-one-out evaluation, FedCHDP achieves the best average F1 of 38.10±11.27, outperforming IVLP(33.90±13.31), MaPLe(35.51±14.11), and VPT(32.69±11.16), confirming improved cross-site generalization.

Computation&Comm Cost(R2): FedCHDP optimizes only 0.45MB(0.36% of CLIP’s 124.5 MB).Prototype exchange adds a negligible 16 KB/round. This is lighter than MaPLe(1.53MB) and substantially reduces communication cost compared with full fine-tuning.



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.

    While the paper has clear strengths in problem motivation and empirical results, Reviewer #1 raises substantial concerns about clarity, notation, and prototype aggregation that currently make the method difficult to follow. Therefore, I invite the authors to submit a rebuttal. The rebuttal should primarily address: (1) clarifying the notation and method pipeline, (2) justifying or improving the simple prototype aggregation strategy, (3) discussing the reliance on manual concept annotations and potential robustness, and (4) adding missing experimental details such as variance reporting and communication overhead.

  • 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 reviewing the authors’ rebuttal and the post-rebuttal feedback from all reviewers, I recommend Accept. The authors have convincingly addressed the major concerns raised during the initial review, including clarity of notation and pipeline, prototype aggregation strategy, fairness of baseline comparisons, and missing experimental details.



Meta-review #2

  • 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.

    Strengths: The paper presents a clear and clinically meaningful motivation by introducing a federated prompt-learning framework (FedCHDP) for medical image classification. The reviewers acknowledge the strength of the empirical results and the value of leveraging multimodal capabilities under federated constraints. In the rebuttal, the authors successfully clarified several notation ambiguities and provided valuable missing experimental details, including variance reporting and communication overhead. Weaknesses and Areas for Improvement: Despite these positive aspects and the authors’ active effort during the rebuttal phase, critical methodological concerns raised in the meta-review regarding the prototype aggregation strategy and concept reliance have not been rigorously or empirically addressed: 1.Inadequate Justification of the Prototype Aggregation Strategy: A core concern was the simple, sample-size-weighted prototype aggregation strategy, which could be highly sub-optimal under severe client data imbalance. In the rebuttal, the authors offered a purely qualitative defense, arguing that sample-size weighting reduces the influence of high-variance local subsets and pointing out that prototypes are only used for prompt regularization. However, this explanation lacks empirical backing. To convincingly justify or improve this design choice, the authors should have provided comparative experiments contrasting their strategy against standard alternative aggregation methods (e.g., class-balanced aggregation or attention-based fusion) to demonstrate its actual empirical validity. 2.Insufficient Analysis of Reliance on Manual Concept Annotations: The framework relies heavily on manual, clinician-annotated clinical concepts during the training phase. While the authors clarified that these annotations are not required during inference, the dependency on expensive human labels during training remains a heavy bottleneck. The authors argued that an uncertainty-aware gate helps mitigate noisy predictions, but they failed to provide concrete quantitative evaluations to prove this robustness. To truly validate the method’s practicality and robustness, the authors should have conducted experiments using concepts generated by Large Language Models (LLMs) or automated pipelines as a proxy to evaluate how performance degrades under varying degrees of annotation noise or in the absence of expert annotators. Summary: In conclusion, while the paper targets an important problem and demonstrates competitive performance, the rebuttal fails to provide the necessary empirical depth to resolve the core algorithmic and robustness concerns. The lack of concrete experimental comparisons for the prototype aggregation choices and the unverified robustness against concept annotation costs prevent the paper from acceptance in its current form. Therefore, a rejection is recommended. The authors are strongly encouraged to incorporate rigorous comparative experiments regarding these components in a future revision.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    The initial reviews were not fully aligned. R2 and R3 considered the paper marginally above the acceptance threshold, mainly due to the clinical relevance, coherent framework design, and strong empirical results. R1, however, raised substantial concerns about method clarity, notation, the definition of visual prompts and semantic anchors, the simplicity of sample-size-weighted prototype aggregation, and the dependence on manual concept annotations. The initial meta-review therefore correctly focused the rebuttal on these points. The rebuttal improved the assessment. It clarified the notation for concept labels and concept values, explained the role of semantic anchors and learnable concept-library embeddings, clarified that classification is performed by image-text matching rather than by global prototypes, and specified that global prototypes are used for contrastive visual alignment. These clarifications address the most serious readability and pipeline concerns raised by R1 and R2.The authors also provided additional information on label heterogeneity, variance across seeds, stronger comparisons, leave-one-site-out evaluation, and communication/computation cost. While such additional numerical results cannot be treated as a substitute for evidence that was absent from the original submission, they help contextualize the reviewers’ concerns and support the plausibility of the submitted claims. Some limitations remain. The technical novelty is moderate, since the method combines concept-guided prompting, uncertainty-aware fusion, and prototype-based alignment rather than introducing a fundamentally new learning principle. The reliance on clinician-annotated concept labels may limit scalability, even if the authors clarified that such annotations are only required during training. The sample-size-weighted prototype aggregation remains a relatively simple strategy and may be suboptimal under more extreme client imbalance or heterogeneous data quality. The presentation should also be revised carefully, as several reviewers noted unclear notation, insufficient figure annotation, and minor typographical issues. Overall, after weighing the original paper, the initial reviews, and the rebuttal, the positive aspects are sufficient for acceptance.



back to top