Abstract

Medical Visual Question Answering (Med-VQA) is pivotal for assisting clinical decision-making and automated diagnosis. While deep learning has achieved remarkable success in general VQA tasks, Med-VQA research targeting specific diseases—particularly pediatric myopia involving multimodal data—remains largely underexplored, primarily due to the scarcity of publicly available, expert-annotated datasets. In this paper, we present the Pediatric Myopia VQA dataset (PM-VQA). To the best of our knowledge, this is the first publicly available Med-VQA dataset dedicated to pediatric myopia. It comprises 2,741 multimodal images (CFP and OCT), totaling 14,781 expert-verified question-answer pairs(QA pairs). Furthermore, we propose an enhanced framework named Proto-CAP, designed to address the semantic gap and overfitting issues inherent in medical small-scale data. Within Proto-CAP, we introduce a Question-Guided Prototype Fusion mechanism and a Curriculum-Adaptive Poly-2 Loss to enhance prior knowledge injection and dynamically balance training difficulty. Experiments on PM-VQA demonstrate that our proposed method achieves state-of-the-art performance, reaching an accuracy of 79.19% and an F1-score of 78.74%, which outperforms existing methods by 2.10% in accuracy and 2.14% in F1-score.Code and dataset will be available at https://github.com/PixelOptic-Dev/Proto-CAP.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3429_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/PixelOptic-Dev/Proto-CAP

Link to the Dataset(s)

PM-VQA: https://pan.baidu.com/s/1BOsD3Tl5uHCuJrUWKLoajQ?pwd=9c35

BibTex

@InProceedings{YanXu_ProtoCAP_MICCAI2026,
        author = { Yan, Xu AND Wu, Yanlin AND Wang, Zixun AND Yu, Jingtao AND Wei, Ruihua AND Li, Tao},
        title = { { Proto-CAP: Prototype Memory Fusion and Curriculum-Adaptive Loss for Pediatric Myopia Visual Question Answering } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses medical visual question answering for pediatric myopia, which is a clinically relevant but relatively underexplored setting. The main contribution is two-fold. First, the authors introduce PM-VQA, a multimodal pediatric myopia VQA dataset built from clinical data and expert-verified question-answer pairs. Second, they propose Proto-CAP, an MMQL-based framework that augments visual representations with question-guided prototype memory fusion and uses a curriculum-adaptive Poly-2 loss to adjust optimization emphasis during training. The paper reports improved performance over several Med-VQA baselines on the proposed dataset and provides ablation results supporting the contribution of both proposed components.

  • 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 studies a clinically meaningful problem. Pediatric myopia is important in practice, and the Med-VQA setting is relevant and well-motivated.
    2. The dataset effort is valuable. PM-VQA appears to be a carefully curated multimodal dataset with expert-verified QA pairs collected from clinical data.
    3. The proposed method is reasonably matched to the task. Both the prototype-based fusion and the adaptive loss are motivated by the challenges of small-scale medical VQA.
    4. The paper includes comparisons with several baselines, ablation studies, and qualitative examples.
    5. The writing is generally clear. The overall pipeline is clear, and the qualitative visualizations help illustrate the intuition behind the method.
  • 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. My main concern is the data split protocol. The paper says the data are split into train/validate/test QA pairs, but it is unclear whether the split is image-based or patient-based. Since multiple QA pairs may come from the same image, this should be clarified carefully.
    2. Table 2 needs stronger justification. Some compared methods are not sufficiently introduced, and it is therefore unclear why these baselines were selected. It would also strengthen the paper to compare with a more recent Med-VQA method, or at least discuss the difference more explicitly.
    3. There are two issues about Table 3.First, why are Accuracy and Recall exactly the same for every row? If this is expected due to the metric definition, please explain it clearly. Second, Precision drops when either Proto or CAP is added alone, but improves when both are combined. The paper should give a clearer explanation of this interaction.
    4. The contribution of the proposed loss needs to be clarified further. Its formulation appears closely related to PolyLoss [1], but the current presentation does not make it clear what the substantive difference is and why the proposed variant is necessary in this setting.
    5. The paper would benefit from more details on the dataset and implementation. For example, it would be helpful to report the number of patients, the distribution of question types, whether all samples contain all modalities, and several missing training details needed to fully understand the setup. [1]. Leng, Z. , Tan, M. , Liu, C. , Cubuk, E. D. , Shi, X. , Cheng, S. , & Anguelov, D. (2022). PolyLoss: A polynomial expansion perspective of classification loss functions [Preprint]. arXiv. doi:10.48550/arXiv. 2204.12511
  • 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 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?

    This paper has several strengths, including a clinically relevant problem setting, a potentially useful pediatric myopia VQA dataset, and a method that shows consistent gains over multiple baselines. In particular, the combination of a clinically grounded multimodal dataset and a task-specific Med-VQA framework makes the work potentially valuable to the community. ButI still have some important concerns. The main issue is that the evaluation protocol is not fully convincing in its current form, since the paper states that the data were split into training, validation, and testing QA pairs, but does not clearly establish an image-disjoint or patient-disjoint split. Because multiple QA pairs can come from the same image, there is a possible leakage concern and it should be clarified. In addition, the methodological novelty is moderate, the validation is limited to a single internal dataset, and several details relevant to reproducibility are missing. Overall, I think the strengths slightly outweigh the weaknesses at this stage, and I am therefore placing the paper marginally above the acceptance threshold, while still expecting the authors to clarify the evaluation protocol and strengthen the presentation in rebuttal.

  • 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 rebuttal addressed most of my concerns.



Review #2

  • Please describe the contribution of the paper

    The paper makes two main contributions. First, it introduces PM-VQA, which is presented as the first expert-verified multimodal pediatric myopia Med-VQA dataset, built from CFP, OCT, and clinical records. This is potentially valuable because disease-specific ophthalmic VQA benchmarks are still scarce.

    Second, the paper proposes Proto-CAP, an extension of MMQL with a Prototype Memory Fusion module and a Curriculum-Adaptive Poly-2 Loss. The method aims to improve semantic alignment and reduce overfitting in small-scale medical VQA. Overall, the dataset contribution is, in my opinion, the more important part of the paper.

  • 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.

    Important application setting. Pediatric myopia is clinically relevant, and Med-VQA for this disease is underexplored. The paper addresses a meaningful real-world problem.

    Potentially useful dataset contribution. PM-VQA fills a gap in multimodal ophthalmic Med-VQA. If publicly released with clear protocols, it could become a useful benchmark for future work.

    Reasonable method design. The prototype fusion and adaptive loss are both well motivated for small-scale medical data, and the method is easy to understand.

    Clear empirical presentation. The paper includes baseline comparisons, ablations, and qualitative examples, which make the proposed improvements easy to follow.

  • 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 novelty is limited. The proposed Prototype Memory Fusion mainly combines prototype/memory learning, question-guided attention, and gated fusion, which are all familiar ideas in vision-language and medical multimodal learning. The paper does not clearly explain what is fundamentally new beyond adapting these components to this task. A similar concern applies to the Curriculum-Adaptive Poly-2 Loss. Dynamic reweighting, curriculum learning, and hard-sample emphasis have all been widely studied, so the current loss design appears incremental rather than conceptually novel.

    The paper should better position itself relative to prior work on memory/prototype learning and adaptive losses. Relevant references include Bengio et al. (2009), Santoro et al. (2016), Snell et al. (2017), Sukhbaatar et al. (2015), Lin et al. (2017), and Leng et al. (2022).

    The experimental validation is still limited. The reported gains over MMQL are promising but relatively modest, and all results are shown only on the newly introduced PM-VQA dataset. There is no external validation on another Med-VQA benchmark. The paper also does not report variance, confidence intervals, or repeated runs with different random seeds. On a relatively small medical dataset, such reporting is important to show that the improvements are stable.

    In addition, comparisons are mainly against prior Med-VQA baselines, but there is limited discussion of stronger modern multimodal baselines or pretrained vision-language models. This makes the “state-of-the-art” claim somewhat less convincing.

    The paper’s value depends heavily on whether the dataset will be released. In my view, the dataset is the most impactful part of the submission. However, the paper is not sufficiently clear about whether PM-VQA will be publicly released, and if so, under what access conditions.

  • 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?

    This paper studies an important and underexplored problem, and I appreciate the effort to build a pediatric myopia multimodal Med-VQA dataset. The method is reasonable, the paper is clearly written, and the results are promising.

    However, I am less convinced by the level of technical novelty, as both the prototype-based fusion and adaptive loss seem to be incremental extensions of existing ideas. The experimental validation is also somewhat limited, since it is restricted to the new dataset and lacks stronger robustness analysis.

    Overall, my recommendation is cautious at this stage. That said, if the authors commit to publicly releasing PM-VQA with the benchmark split and evaluation code, I would be willing to raise my score to weak accept, because the dataset contribution would then provide clear community value.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper makes three contributions to medical visual question answering (Med-VQA) targeting pediatric myopia: 1- PM-VQA dataset. The authors construct what they claim to be the first publicly-oriented, expert-verified Med-VQA dataset dedicated to pediatric myopia. It comprises 2,741 multimodal images (Color Fundus Photography and Optical Coherence Tomography) collected from a collaborating hospital, with 14,781 question–answer pairs formulated by two ophthalmologists and validated by a senior expert. The data is partitioned into 10,271 / 2,287 / 2,223 train/val/test QA pairs. 2- Prototype Memory Fusion (Proto). A learnable memory bank M ∈ R^{K×D} of medical concept prototypes is queried by a global text representation (mean-pooled BERT embeddings) via attention; the retrieved concept vector C is fused with Swin-Transformer visual features V through a learned sigmoid gate, V_enhanced = LayerNorm(g·V + (1−g)·C). The aim is to inject domain priors and bridge the visual–language semantic gap. 3- Curriculum-Adaptive Poly-2 Loss (CAP). An extension of Poly-Loss (Leng et al., 2022) where the polynomial coefficients e₁ and e₂ are no longer static but scheduled with respect to current training accuracy: e₁ follows a cosine decay (fast early convergence) while e₂ increases linearly (late-stage hard-example mining). The aim is self-paced curriculum learning. The two modules are integrated on top of the MMQL multi-question backbone (Chen et al., 2024), and their combination yields 79.19% accuracy and 78.74% F1 on PM-VQA, +2.10 / +2.14 points over MMQL.

  • 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- Pediatric myopia is a major public-health issue (cited prevalence of ~50% of the world’s population by 2050). Existing Med-VQA benchmarks (VQA-RAD [7], SLAKE [12]) are radiology-centric, while existing ophthalmology VQA datasets (DME-VQA [19], INSEGCAT-VQA [20], OphthalWeChat [22]) are largely repurposed from public datasets, single-modality, or lack QA verification (Table 1). PM-VQA is the only entry that is simultaneously clinical-source, multimodal (CFP+OCT) and QA-verified, which is a real value-add for the community. 2- QA pairs are formulated by two ophthalmologists and re-validated by a senior expert, with explicit quality control and de-identification stages (Fig. 1). Ethics committee approval is reported. This is stronger annotation governance than most ophthalmology VQA datasets. 3- The Prototype Memory Fusion is a sensible inductive bias for small-scale medical data: parameterising medical concepts as a learnable memory addresses the well-known data-scarcity problem cited in the introduction. The gating mechanism preserves the original visual stream, mitigating the risk that the prototype path overrides image evidence — a nice design choice. The Curriculum-Adaptive Poly-2 Loss is a clean, lightweight extension of Poly-Loss [8] that ties coefficient scheduling to a model-state signal (accuracy). The cosine-decay / linear-increase schedule for e₁ / e₂ is consistent with the easy-to-hard curriculum intuition. 4- Seven Med-VQA baselines are compared (CMSA, M-Mixup, M2I2, MUMC, MEVF-BAN, Surgical-VQA, MMQL). The improvement over the closest baseline (MMQL) is non-trivial (+2.10% accuracy, +2.14% F1). 5- Table 3 ablates Proto and CAP independently and jointly, supporting the claim that the two modules are complementary (each adds something on its own; the combination is super-additive in F1: +0.98 + +0.74 individually vs +2.14 jointly). 6- The qualitative attention-map comparison in Fig. 3 (right panel) shows visibly more focal attention for Proto-CAP than for MMQL, lending some plausibility to the ‘medical-prior injection’ interpretation.

  • 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- Each component is a thin variation on prior art:

    • The Prototype Memory Fusion is structurally a memory-augmented attention bank with a sigmoid gate. Memory networks for medical image analysis and gated text-conditioned visual fusion are well-established (e.g., dynamic memory networks for VQA, gated multimodal units of Arevalo et al. 2017, conditional reasoning of MEVF-BAN [23] which is itself a baseline here). The paper does not articulate what is fundamentally new beyond combining a learnable memory bank with a BERT-pooled query and a sigmoid gate.
    • The Curriculum-Adaptive Poly-2 Loss is a parameter-scheduling modification of Poly-Loss [8]. The schedules for e₁ (cosine decay) and e₂ (linear increase) are not formally specified (no equations, no hyperparameters, no sensitivity analysis). Tying loss weights to training accuracy is a small step beyond [8], and self-paced / curriculum-driven loss-shaping has a long history (Kumar et al. 2010; Bengio et al. 2009; focal-loss family). Without an explicit comparison against vanilla Poly-Loss with tuned static (e₁,e₂), the added value of the adaptive aspect is not isolated.
    • Overall the framework is MMQL [1] + memory bank + scheduled Poly-Loss; the +2.10% gain is real but modest and may largely come from the prototype module rather than from a conceptual breakthrough. 2- Table 1 advertises PM-VQA as multimodal (CFP and OCT), and the introduction motivates the work by the need for joint CFP/OCT analysis. However, Section 3 describes only a single image I processed by one Swin Transformer, with no explanation of how CFP and OCT are jointly fed to the model, aligned per-patient, or fused across modalities. Are CFP and OCT treated as independent samples (in which case the model is monomodal at inference time)? Are they concatenated, encoded by separate branches, or paired per patient? This is the most important methodological gap of the paper given that multimodality is an explicit selling point. Without clarification, the ‘multimodal’ claim is not supported by the method. 3- All experiments are run on the authors’ own PM-VQA. There is no evaluation on existing ophthalmology VQA datasets (DME-VQA, INSEGCAT-VQA, OphthalWeChat) nor on radiology benchmarks (VQA-RAD, SLAKE) to demonstrate that the Proto and CAP modules generalise. Given the reliance on a domain-specific memory bank, it is particularly important to show whether the gains transfer or whether they are tied to PM-VQA-specific concepts. 4- Tables 2 and 3 report single-point estimates with no standard deviation, no confidence interval, no multi-seed runs and no statistical test. With +2.10% accuracy on a 2,223-QA test set, the naive 95% Wilson CI for a binomial proportion is roughly ±1.7%, so the claimed gain over MMQL is barely outside the noise floor. For the Proto-only variant (+1.18%) and CAP-only variant (+0.65%) the claim is statistically not supported by the evidence shown. 5- The dataset section omits information that is essential to assess scientific value and risk of bias:
    • Number of unique patients (vs. number of images).
    • Patient demographics (age range, sex distribution, ethnicity, spherical-equivalent distribution, axial-length distribution).
    • Question type taxonomy and answer space (open vs. closed; answer-vocabulary size; class imbalance).
    • Inter-annotator agreement between the two ophthalmologists and the senior expert (Cohen’s κ or analogous).
    • Image-level metadata (device manufacturer, resolution, OCT scan pattern), which affects external validity.
    • How CFP and OCT are paired per patient and per QA pair.
    • Train/val/test split protocol is it patient-level (no leakage) or QA-pair-level (likely leakage)? This is critical because many QA pairs share an image; QA-level random splitting would let the model see the same image at train and test time.
  • 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 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.

    (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?
    • PM-VQA is a real and currently missing resource pediatric myopia is a high-impact public-health topic and existing ophthalmic VQA datasets are either monomodal, public-source or unverified. Releasing a clinical, multimodal, expert-verified dataset would be a valuable community contribution.
    • The proposed modules are well-motivated, simple to integrate, and the ablation supports their complementarity.
    • Empirical gains over a strong recent backbone (MMQL [1]) are consistent across all four reported metrics.
    • Methodological novelty is incremental: a memory bank with gated fusion + a scheduled Poly-Loss on top of an existing backbone. None of the three contributions is individually MICCAI-strong; the dataset is the strongest item but its size (2,741 images) and the missing description of the multimodal pipeline limit its standing.
    • The ‘multimodal’ claim central to the paper’s differentiation from existing eye-VQA datasets is not demonstrated in the method or experiments. The model as described in Section 3 is a single-image pipeline.
    • Single-dataset evaluation, no significance testing, no multi-seed variance, and a +0.65/+1.18% per-module gain that is below the binomial-CI noise floor for a 2,223-sample test set, jointly weaken the empirical evidence.
    • Reproducibility is poor: no code, no data link, no commitment to release, and several hyperparameters are unstated.
    • Several clarifications are needed regarding ethics on paediatric data and the rationale for demographic-prediction QA items. Issues that should be addressed in rebuttal:
    • Clarify how CFP and OCT are jointly used (per-patient pairing, encoder topology, fusion stage). If the model is in fact monomodal, soften the multimodal claim.
    • Provide multi-seed mean ± std and a statistical test (McNemar or paired bootstrap) for Tables 2 and 3.- Describe the train/val/test split protocol (patient-level vs QA-level) and confirm absence of image-level leakage.
    • Specify K, p, and the closed-form schedules for e₁ and e₂; include a sensitivity analysis on K.
    • Add at least one cross-dataset experiment (e.g., DME-VQA or VQA-RAD) demonstrating that the Proto and CAP modules are not specific to PM-VQA.
    • Commit explicitly to releasing code, prototypes, and the PM-VQA dataset under a documented licence.
    • Discuss the inclusion of demographic-prediction QA items (e.g., gender) and report performance broken down by question type.
  • 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

Our CAP loss mimics the clinical training process. In the early stage, when accuracy is low, e1 promotes rapid convergence, allowing the model to quickly learn fundamental and typical ophthalmic anatomical structures. The cosine schedule provides slow initial decay, accelerated mid-stage adjustment, and smooth final convergence, preventing premature or overly rapid gradient decay and reducing training instability. In the later stage, as model proficiency improves, e2 increases linearly, suppressing redundant gradients from easy samples while amplifying attention to difficult clinical cases. Unlike PolyLoss, which usually provides only marginal gains on large-scale datasets, our dynamic curriculum plays a critical role in preventing overfitting in small-scale medical data. Proto provides clinically informative high-dimensional features, while CAP offers a dynamic learning path to fully exploit these features. CAP can therefore be regarded as an enhancement to Proto, and their combination significantly outperforms either module alone.




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 introduces a pediatric myopia medical VQA benchmark (PM-VQA) and a Proto-CAP framework that combines prototype memory fusion with a curriculum-adaptive loss to improve multimodal reasoning performance. The work addresses a clinically meaningful and underexplored problem, and the curated, expert-verified dataset together with consistent empirical improvements over prior Med-VQA baselines make it potentially valuable to the community.

    On the other side, the reviewers also raise main concerns. First, the evaluation protocol needs clarification: the current QA-level split raises potential leakage issues, and the absence of multi-seed results, statistical testing, and robustness analysis makes the reported gains hard to judge. Second, the “multimodal” aspect is unclear, as the paper does not explain how CFP and OCT are jointly modeled or fused. Third, the novelty appears limited, with both the prototype memory and adaptive loss being incremental extensions without clear differentiation from prior work. Fourth, the empirical validation is somewhat limited: it relies on a single internal dataset, includes insufficient comparison to stronger or more recent baselines, lacks robustness and statistical analysis, and does not demonstrate whether the proposed modules generalize beyond PM-VQA. Finally, key dataset and reproducibility details are missing, weakening transparency and rigor.

  • 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.

    This paper proposes Proto-CAP, a pediatric myopia Med-VQA framework with prototype memory fusion and curriculum-adaptive loss, together with PM-VQA, an expert-verified pediatric myopia dataset from CFP and OCT images. The main strength of the work lies in the dataset contribution: pediatric myopia Med-VQA is clinically meaningful and underexplored, and PM-VQA could provide a valuable community benchmark. The rebuttal strengthens this aspect by committing to release the dataset and code upon acceptance. However, several concerns remain only partially resolved. The methodological contribution appears moderate, with prototype memory fusion and curriculum-adaptive loss representing incremental extensions of existing ideas, while validation is limited to a single internal benchmark with modest gains. In addition, the rebuttal clarification weakens the multimodal framing, as CFP and OCT are treated as separate Med-VQA tasks rather than jointly modeled at inference. Overall, my assessment is mixed: with the promised dataset release, I lean toward acceptance because the potential community value outweighs the remaining concerns. However, the methodological contribution alone would not, in my view, be sufficient without public release of the dataset, and I strongly encourage the authors to follow through on this commitment.



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.

    After reviewing the manuscript, the reviews, and the authors’ rebuttal, I recommend acceptance.

    The paper addresses an important and underexplored problem in pediatric myopia assessment and introduces a clinically meaningful multimodal Med-VQA dataset. While the methodological novelty is moderate, the combination of a new benchmark, a well-motivated task setting, and consistent empirical improvements provides a valuable contribution to the community.

    The main concerns raised during review related to the evaluation protocol, potential data leakage, reproducibility, and the scope of validation. In my view, the rebuttal adequately clarifies these issues and strengthens confidence in the reported results. This is reflected in the updated recommendations from Reviewers 1 and 2. Overall, I believe the dataset contribution, clinical relevance, and experimental results outweigh the remaining limitations. The paper represents a useful contribution and I therefore recommend 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.

    Accept

  • Please justify your recommendation.

    The paper addresses a clinically relevant and underexplored problem in pediatric myopia VQA, introducing the PM-VQA dataset and the Proto-CAP framework. The dataset appears to be a valuable contribution, with expert-verified annotations and multimodal ophthalmic data. The proposed method demonstrates consistent improvements over established Med-VQA baselines, and the rebuttal satisfactorily clarifies several key concerns, including confirmation of patient-level data splitting to prevent leakage, justification of baseline selection, clarification of multimodal dataset construction, and additional details regarding the prototype memory and curriculum-adaptive loss design. While the methodological novelty remains somewhat incremental and external validation is limited, the reviewers generally agree that the dataset contribution is significant and that the reported gains are meaningful in this specialized setting. Given the positive reviews, satisfactory rebuttal, and community value of the benchmark, I recommend the paper for acceptance.



back to top