List of Papers Browse by Subject Areas Author List
Abstract
Postoperative Best Corrected Visual Acuity (BCVA) prediction is essential for cataract surgical planning. Existing deep learning methods commonly rely on symmetric fusion, implicitly treating all modalities as equally reliable. However, in clinical practice, auxiliary modalities (e.g., SLO, text) are often noisy or missing, while Optical Coherence Tomography (OCT) provides high-resolution structural evidence of retinal status. Symmetric fusion may distort high-fidelity OCT representations. To address this, we propose Asym-MSF, an asymmetric framework that prioritizes OCT as the foundational anchor. We introduce an Asymmetric Latent Bottleneck (ALB), which generates spatial-aware seeds from OCT to selectively compress auxiliary modalities, effectively filtering out irrelevant noise before fusion. Furthermore, we design a Multi-scale HyperNetwork. Unlike standard MLP-based generators, our module captures multi-granular structural contexts to generate position-specific kernels. To preserve anatomical fidelity during dynamic modulation, we integrate a Self-supervised Feature Restorer (SFR), which reconstructs OCT features under hybrid amplitude, directional, spatial-gradient, and frequency-consistency constraints. Experiments on a clinical dataset of 2,635 cases demonstrate that Asym-MSF achieves the best performance among the compared methods, with a Mean Absolute Error (MAE) of 0.0378. Code is available at https://github.com/wyf-cyber/Asym-MSF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4326_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/wyf-cyber/Asym-MSF
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYif_AsymMSF_MICCAI2026,
author = { Wang, Yifei AND Zhou, Qian AND Han, Xu AND Zou, Hua AND Jiang, Haifeng AND Wang, Yong AND Cao, Danmin},
title = { { Asym-MSF: Asymmetric Multi-scale Structural Fusion for Robust BCVA Prediction with Self-supervised Fidelity Constraints } },
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 addresses the clinically important task of predicting postoperative best-corrected visual acuity (BCVA) for cataract patients using multimodal preoperative data, with particular attention to realistic clinical scenarios in which auxiliary modalities may be noisy or incomplete. The primary contribution is Asym-MSF, an asymmetric multimodal fusion framework that explicitly prioritizes optical coherence tomography (OCT) as the foundational structural modality while treating supplementary modalities such as scanning laser ophthalmoscopy (SLO) and clinical text as contextual support rather than co-equal inputs. The proposed architecture integrates three key innovations: an Asymmetric Latent Bottleneck (ALB) that generates latent queries derived from OCT to compress auxiliary modalities in an OCT-guided manner; a Dynamic Convolutional Modulation (DCM) module featuring a multi-scale kernel generator with spectral gating to inject auxiliary context into the OCT processing stream; and a Self-supervised Feature Restorer (SFR) that reconstructs OCT features from the generated kernels to regularize the modulation process and preserve anatomical fidelity. The authors argue that this structural prioritization aligns with clinical practice in ophthalmology, where OCT provides the highest-fidelity anatomical evidence while SLO and textual data serve as supplementary information that may be incomplete or inconsistent. Experimental results on a clinical dataset demonstrate that the proposed model achieves a lower mean absolute error (MAE) than existing methods, and ablation studies confirm that both the asymmetric design philosophy and the self-supervised restoration objective contribute meaningfully to performance improvements.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has several strengths. First, the core motivation is clinically reasonable and well-matched to the application. Treating OCT as the primary structural modality, while using SLO and text as auxiliary context, is a sensible assumption for BCVA prediction and provides the method with a clear design rationale.
Second, the proposed framework is not just a straightforward fusion variant. The asymmetric latent bottleneck uses OCT-derived queries to compress auxiliary information in a targeted way, which is more thoughtful than standard symmetric fusion. The dynamic modulation design further enhances the conditioning’s adaptability and task specificity.
Third, the self-supervised restoration branch is a meaningful component. It aims to regularize the modulation process so that auxiliary information does not distort the underlying OCT structure, a relevant consideration in medical imaging settings.
Fourth, the empirical results on the reported dataset are promising. The method improves MAE over the compared baselines, and the ablation results are generally consistent with the paper’s main narrative. In particular, comparing different foundational modalities supports the choice of OCT as the anchor modality.
Finally, the paper is fairly easy to follow at a high level. The main idea is clear, the overall framework is intuitive, and the anonymized code link is a positive sign for reproducibility.
- 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 main weakness of the paper is that the experimental evidence is still insufficient to support the scope of its claims. The method is evaluated on a single clinical dataset using a random 8:1:1 split, without external validation, cross-center testing, or clarification on whether the split is performed at the patient level. Since the dataset contains 2,141 eyes from 1,960 patients, this detail is important, as non-patient-level splitting could lead to overly optimistic results. A second concern is that the paper strongly emphasizes robustness to missing modalities, yet this is not validated in a sufficiently direct way. Although the manuscript states that SLO is missing in about 63% of cases, it does not provide controlled experiments under different missingness settings, stratified analyses by missingness pattern, or comparisons between complete and incomplete cases. It is therefore difficult to determine whether the gains truly come from better handling of missing or noisy auxiliary data. The paper also lacks computational analysis. Given the complexity of the proposed framework, including the asymmetric bottleneck, dynamic kernel generation, spectral gating, and restoration branch, it would be important to report parameter counts, FLOPs, runtime, or memory usage. Without such analysis, the practical cost of the method remains unclear. In addition, the method description is not fully self-contained. Some important implementation details are either only briefly mentioned or deferred to prior work, which leaves gaps in reproducibility. The related-work discussion could also be stronger, particularly in positioning the method against broader literature on missing-modality learning and reliability-aware multimodal fusion. There are also several presentation issues, including inconsistent notation in Fig. 1 and inconsistent acronym expansions in Table 2, which reduce clarity. Finally, the evaluation is somewhat limited, with Accuracy quickly saturating and no variance estimates or statistical testing reported. Overall, the idea is interesting, and the results are promising, but the current evidence is not yet sufficient to fully justify the strength of the claims.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
The paper presents a promising and clinically sensible approach, and the asymmetric fusion perspective represents a valuable direction for multimodal medical AI research. However, I believe the most critical revisions should focus not on introducing additional architectural novelty, but on substantially strengthening the empirical evidence supporting the claims. Specifically, the manuscript would benefit significantly from clarification of the patient-level data split methodology and, ideally, more robust validation beyond a single random partition to ensure generalizability. Given that robustness to missing or noisy auxiliary inputs forms a central motivation for the work, dedicated experiments under controlled missing-modality scenarios would provide crucial validation of the framework’s practical utility. Additionally, a comprehensive efficiency analysis—including parameter counts, FLOPs, inference runtime, and memory consumption—is essential to contextualize performance gains relative to computational costs, particularly given the framework’s architectural complexity. The reproducibility of the work would be enhanced by providing more complete implementation details for all modules and training configurations. Furthermore, the related work discussion requires more rigorous positioning, particularly regarding existing literature on missing-modality handling and reliability-aware multimodal learning strategies. Beyond these substantive concerns, I encourage the authors to carefully polish the figures and tables, as several inconsistencies in notation and acronym expansion currently detract from clarity. Overall, while I find the research direction genuinely interesting and potentially impactful, I do not yet believe the empirical support is fully commensurate with the strength of the paper’s claims, and addressing these methodological and presentation issues would substantially strengthen the contribution.
- 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?
I arrived at a weak reject recommendation because the paper has a clear idea and promising results, but the empirical and methodological support is not yet sufficiently complete for acceptance in its current form. On the positive side, the central premise is well motivated, the architectural direction is interesting, and the reported performance on the dataset is strong. The anchor-selection ablation is also aligned with the paper’s core claim and strengthens the narrative.
However, several issues prevent me from being fully convinced. The evaluation is limited to one dataset and one random split, with no external validation and no clear statement about patient-level partitioning, despite the dataset containing multiple eyes from fewer patients than the number of eyes. The paper also strongly emphasizes robustness to missing modalities, yet does not present controlled robustness experiments to directly verify this claim. In addition, the proposed architecture is comparatively heavy, but there is no analysis of computational cost or practical trade-offs. Finally, the manuscript still has several gaps in presentation and reproducibility, including incomplete methodological detail, inconsistent terminology, and somewhat overstated claims.
Overall, I think the paper is promising and could move into the accept range with a convincing rebuttal and stronger evidence, but in its current form, I place it slightly below the bar.
- 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 revise my recommendation to weak accept. The response addresses several of my main concerns and provides useful clarifications about the method, experimental setting, and interpretation of the results. In particular, the rebuttal makes the contribution clearer and alleviates some of my earlier doubts about the validity of the approach. Although some limitations remain and the paper could still benefit from stronger empirical evidence and clearer presentation, I now think the core idea is sound and the results are sufficiently promising. Overall, I believe the paper meets the acceptance bar, albeit somewhat marginally.
Review #2
- Please describe the contribution of the paper
This work propose an asymmetric multimodal fusion framework for postoperative BCVA prediction, which explicitly treats OCT as the structural anchor and uses auxiliary modalities only for adaptive refinement. This paper design an Asymmetric Latent Bottleneck, a multi-scale dynamic convolutional modulation module, and a self-supervised feature restorer to achieve noise-robust fusion while preserving anatomical fidelity. The authors emphasize that extensive experiments on a clinical cataract dataset demonstrate that the proposed method achieves state-of-the-art performance and validates the importance of asymmetric modality prioritization.
- 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.
Well-motivated clinical problem formulation：The paper identifies a meaningful limitation of prior multimodal BCVA prediction methods, namely the use of symmetric fusion despite an inherent hierarchy among modalities. The argument that OCT should be treated as the structural anchor, while SLO and text should play supporting roles, is clinically intuitive. The proposed method achieves the best performance among all compared methods, outperforming both the strongest baseline UML and the DyKen-Hyena architecture.
- 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.
While the overall framework is well assembled, several components are conceptually related to existing ideas: the bottlenecked cross-attention resembles Perceiver-style compression, the dynamic kernel generation builds upon prior dynamic-kernel formulations, and the self-supervised restorer uses reconstruction-based regularization with gradient/frequency constraints. As a result, the main novelty seems to lie more in the asymmetric clinical formulation and the integration of existing ideas than in a fundamentally new modeling primitive. Limited validation scope. The authors emphasize that the method preserves anatomical fidelity and filters noisy auxiliary signals, but these claims are mostly supported indirectly through performance improvements.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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 has solid motivation, a consistent framework, and competitive empirical performance, but its impact is somewhat limited by the lack of broader validation and by novelty that is more integrative than fundamentally methodological.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes an asymmetric multimodal fusion framework (Asym-MSF) for BCVA prediction, which prioritizes OCT as the structural anchor to address noise and missing data in auxiliary modalities. It introduces an Asymmetric Latent Bottleneck to selectively filter auxiliary information, a multi-scale HyperNetwork for dynamic structural modeling, and a self-supervised Feature Restorer to preserve anatomical fidelity. Experiments on a clinical dataset demonstrate improved performance and robustness over existing methods.
- 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 introduces a well-motivated asymmetric fusion paradigm that aligns with clinical priors, effectively addressing noise contamination and modality incompleteness by prioritizing OCT as a structural anchor. 2.The proposed multi-scale dynamic modulation mechanism is technically sound and capable of capturing complex anatomical structures at different granularities, improving representation quality. 3.The integration of self-supervised fidelity constraints helps preserve structural consistency and leads to strong empirical performance on real clinical data.
- 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.I suggest the authors ensure that the symbols used in the figures are consistent with those in the textual descriptions, as this would improve clarity and make the method easier to follow.
2.The paper claims to address the issue that symmetric fusion may contaminate pristine OCT features with auxiliary noise. However, this problem appears to be directly handled only by the SFR module. I recommend revising this statement for better alignment with the method. Additionally, the ablation study lacks results for using the SFR module alone.
3.In the methodology section, it would be helpful to explicitly state that Asym-MSF is designed for Best Corrected Visual Acuity (BCVA) prediction, which would improve overall readability and understanding.
I look forward to seeing an extended journal version of this work
- Please rate the clarity and organization of this paper
Poor
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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.
(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?
This paper tackles a practical problem in multimodal medical prediction under missing or noisy modalities. The proposed asymmetric fusion framework is well-motivated and technically sound, with reasonable novelty in its design. Experimental results on a real clinical dataset demonstrate improved performance and robustness over prior methods.
However, some issues remain, including insufficient clarity in linking the motivation to certain components, limited ablation studies, and minor presentation problems.
Overall, I recommend a accept.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The author resolved my main questions, and I will maintain a positive attitude.
Author Feedback
We thank the reviewers and Meta-Reviewer for their constructive feedback. Q1: Missing/noisy auxiliary modalities. (R1, Meta) A1: We study real clinical missingness: OCT/text are routinely available, while SLO is absent in ~63% of cases. Asym-MSF therefore anchors on OCT and uses SLO/text as auxiliary context. It improves over UML in this setting (MAE 0.0457→0.0378; Acc 96.25%→98.60%). Controlled dropout, which artificially masks auxiliary modalities in complete cases to simulate missingness rates, was not used as main evidence because it would rely on the smaller complete-modality subset and may introduce selection bias. We will revise the claim to real-world SLO missingness, not exhaustive robustness to all missingness patterns. Q2: Validation scope and split protocol. (R1, Meta) A2: Only 181 eyes (8.5%) in our dataset share patient identity with another eye, so patient-level overlap is limited. This eye-level protocol is also consistent with prior BCVA prediction studies [1,2]. External validation was omitted because no comparable clinical dataset with the same multimodal inputs and postoperative BCVA labels is available. All baselines and Asym-MSF used the same split, preprocessing, and metrics, so the relative gains are unlikely to arise from protocol differences. We will clarify the protocol and avoid overstating patient-independent or cross-center generalization. Q3: Computational cost. (R1, Meta) A3: Asym-MSF has 34.19M training parameters and 28.88M active inference parameters, since SFR is training-only, compared with 27.20M for UML. Its inference complexity is 5.17 GFLOPs/sample versus 4.40 for UML, a 17.6% increase. Although moderately heavier, Asym-MSF improves MAE from 0.0457 to 0.0378 and Acc from 96.25% to 98.60% over UML. Since postoperative BCVA prediction is not a real-time task, this trade-off is practical. We will report these implementation-derived statistics and clarify the training/inference distinction. Q4: Method clarity and presentation. (R1, R3) A4: We will improve the self-contained description of ALB/DCM/SFR, explicitly state that Asym-MSF is designed for postoperative BCVA prediction, and fix notation/acronym inconsistencies in figures and tables. Q5: SFR role and ablation. (R3) A5: We will clarify that auxiliary-noise mitigation comes from ALB, DCM, and SFR jointly, not SFR alone. ALB filters auxiliary evidence with OCT-derived queries, DCM injects it via controlled modulation, and SFR regularizes DCM-generated kernels to preserve OCT anatomical fidelity. Because SFR depends on DCM-generated kernels, an “SFR-only” setting is not a valid standalone model; therefore, we evaluate SFR by comparing the full model with its no-SFR variant. Q6: Novelty and relation to existing ideas. (R2) A6: Our contribution is an OCT-anchored asymmetric fusion design for BCVA prediction, using reliable OCT structure to guide potentially incomplete/noisy SLO-text context. ALB differs from [3] by generating spatial-aware seeds from OCT features, so SLO/text are compressed according to OCT anatomy rather than modality-agnostic tokens. DCM differs from [4] by replacing the MLP-based kernel generator with multi-branch dilated convolutions to model multi-scale structural patterns. Together with SFR, these modules leverage auxiliary information while preserving OCT structure. Our comparisons and ablations support the contribution of this asymmetric design. [1] Incomplete Multimodal Learning for Visual Acuity Prediction After Cataract Surgery Using Masked Self-Attention. MICCAI (2023). [2] UML: A Unified Multimodal Learning Framework for Cataract Postoperative Visual Acuity Prediction with Uncertain Missing Modalities. ICASSP (2025). [3] Perceiver IO: A general architecture for structured inputs & outputs. ICLR (2022). [4] DyKen-Hyena: Dynamic Kernel Generation via Cross-Modal Attention for Multimodal Intent Recognition. arXiv (2025).
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.
The paper addresses a clinically important problem with a well-motivated asymmetric fusion design. One reviewer recommends Accept (5), one Weak Accept (4), and one Weak Reject (3). The main concerns center on limited validation (single dataset, no external validation, unclear patient-level splitting), lack of controlled experiments for missing-modality robustness, and missing computational cost analysis. These concerns are addressable during rebuttal. Given the strong support from Reviewer 3 and the clinical relevance of the problem, the paper should proceed to rebuttal with a clear expectation that the authors address the validation and robustness concerns.
- 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 a well-motivated asymmetric multimodal fusion framework for BCVA prediction, anchored on OCT with auxiliary SLO/text inputs, which aligns well with clinical practice. Reviewer #3 gave Accept (5), Reviewer #2 gave Weak Accept (4), and Reviewer #1’s Weak Reject (3) raised concerns about validation scope and computational cost—issues that were adequately addressed in the rebuttal. The overall consensus supports acceptance, and the core methodological contribution remains solid.
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.
This paper proposes Asym-MSF, an asymmetric multimodal fusion framework for postoperative BCVA prediction that prioritizes OCT as the structural anchor while incorporating auxiliary modalities for refinement. Reviewers agreed that the problem is clinically important and that the asymmetric fusion design is well motivated. The method demonstrates promising performance improvements on a real clinical dataset.
The main concerns raised during review included limited validation scope (single dataset, no external validation), insufficient direct experiments for missing-modality robustness, lack of computational cost analysis, and several presentation issues. In the rebuttal, the authors clarified the split protocol, contextualized the missing-modality setting using real-world clinical missingness, provided parameter/FLOPs analysis, and better explained the roles of ALB, DCM, and SFR. The rebuttal also clarified the methodological novelty and appropriately moderated several claims.
Importantly, Reviewer #1 updated their score from Weak Reject to Weak Accept after rebuttal, while the other reviewers maintained positive assessments. Overall, the paper presents a clinically meaningful and technically sound contribution with convincing empirical results despite some remaining limitations in validation scope.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After reading the rebuttal and other reviewers’ comments, I think the authors have provided reasonable clarifications to the reviewers’ concerns.