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

Abstract

Multi-modal fundus image analysis holds significant clinical value for the early diagnosis of ophthalmic diseases. While existing studies on multi-modal diagnosis predominantly focus on designing complex fusion architectures, they often overlook the underlying optimization dynamics during joint training. In this work, we reveal that optimization conflicts between the OCT and CFP encoders and the fusion module lead to suboptimal model performance. To address this, we propose a novel framework incorporating Fusion-Guided Gradient Decoupling Learning (FGDL) and a Cross-Modal Semantic Alignment (CMSA) module. By decoupling optimization paths and employing conflict-aware projection to rectify uni-modal gradients, FGDL eliminates inter-modal gradient interference while guiding uni-modal features toward the global multi-modal optimum. Meanwhile, the CMSA module achieves semantic alignment between OCT structural features and CFP texture features by capturing global dependencies. Furthermore, we introduce a loose pairing strategy to enhance the model’s robustness against data heterogeneity. Experiments on the MMC-AMD and GAMMA datasets demonstrate that our method achieves accuracies of 93.01% and 95.00%, respectively, significantly outperforming existing state-of-the-art approaches.Code is available at https://github.com/lxz8763/FGDL-Fundus.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3644_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/lxz8763/FGDL-Fundus

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuXia_Addressing_MICCAI2026,
        author = { Liu, Xiaozhou AND Qiu, Zhike AND Zeng, Luping AND Hu, Jiahui AND Liu, Jiahao AND Wen, Liangming AND Du, Zhiguo},
        title = { { Addressing Gradient Conflicts in Multimodal Fundus Disease Recognition with Fusion-Guided Learning } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper studies multimodal fundus disease recognition using OCT and CFP images. Instead of focusing only on designing a stronger fusion block, the paper tries to address optimization issues during joint multimodal training. To this end, it proposes a framework that combines three components: a loose pairing strategy to increase training diversity, a cross-modal semantic alignment module to enhance feature interaction, and a fusion-guided gradient decoupling learning strategy to reduce conflicts between unimodal and fusion optimization. Experiments on the MMC-AMD and GAMMA datasets show improved classification performance over several unimodal and multimodal 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 paper addresses a relevant and practical problem in multimodal fundus disease recognition. Combining OCT and CFP is clinically meaningful, since the two modalities provide complementary structural and appearance information.

    2.A notable strength of the paper is that it studies multimodal learning from an optimization perspective, rather than only proposing another fusion architecture. This angle is interesting and gives the work a clearer motivation.

    3.The overall framework is reasonably complete. The paper combines optimization design, cross-modal interaction, and a data augmentation strategy into one pipeline, and the roles of the different modules are generally understandable.

    4.The empirical results on the two public datasets are competitive, and the paper includes ablation studies and additional gradient-related analyses beyond reporting only final accuracy. This helps make the method evaluation more informative.

  • 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 optimization idea in FGDL is interesting, but its methodological novelty would benefit from clearer positioning with respect to prior work. In particular, the gradient projection step appears conceptually close to gradient surgery methods such as PCGrad, which also project conflicting gradients to remove interfering components. In addition, the use of stop-gradient to decouple optimization paths is itself a known training mechanism. Therefore, the current paper would be strengthened by clarifying more explicitly what is fundamentally new in FGDL beyond adapting these ideas to the multimodal OCT-CFP setting. Yu T, Kumar S, Gupta A, et al. Gradient surgery for multi-task learning[J]. Advances in neural information processing systems, 2020, 33: 5824-5836. Chen X, He K. Exploring simple siamese representation learning[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021: 15750-15758.
    2. Pairing an OCT image with any CFP image from the same category increases the amount of training data, but it also weakens strict sample-level correspondence between modalities. This may partly change the nature of the task from true paired multimodal learning toward category-level association learning. The clinical realism of the LoosePair strategy would benefit from further discussion.
    3. The cross-modal semantic alignment module appears relatively conventional, as it mainly relies on a combination of intra-modal attention and cross-modal attention for feature interaction. In contrast, the optimization-related design for handling gradient conflict seems to be the more distinctive part of the work. Therefore, emphasizing CMSA as the primary contribution may not be entirely appropriate.
    4. FGDL includes two steps: first, it blocks the fusion branch from directly updating the unimodal encoders; second, it further adjusts the unimodal gradients with projection. Since the first step already removes the main source of interference, it is not clear how much extra benefit the projection step really brings. As a result, the current experiments do not clearly show whether the improvement comes mainly from decoupling, from projection, or from other components such as LoosePair.
    5. Figure 1 mainly shows changes in gradient magnitude and training stability. However, these observations alone are not sufficient to establish the existence of gradient conflict. Gradient conflict is fundamentally a directional notion and is usually characterized by cosine similarity or gradient inner products. Smaller gradients may simply indicate that training is approaching a local optimum, while unstable gradients may also arise from factors such as learning rate choice, batch noise, or data heterogeneity. Therefore, the evidence in Figure 1 is suggestive but not sufficient to rigorously support the claim of optimization conflict.
    6. Figure 3 shows larger gradients after applying FGDL. However, this does not necessarily mean that gradient suppression is resolved. The increase could simply be due to the stop-gradient design, which removes interference from the fusion branch and allows unimodal gradients to become larger.
  • 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.

    (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?

    This paper addresses an important problem in multimodal fundus disease recognition and has a clear practical motivation. A positive aspect is that the paper does not only focus on fusion architecture design, but also considers optimization behavior during joint multimodal training. The reported experimental results are also competitive.

    However, I still have several concerns. The novelty of FGDL is not yet fully clear, since its design appears close to combining known gradient projection ideas with stop-gradient based decoupling. In addition, the contributions of different components are not clearly separated: LoosePair already brings a large gain, while FGDL itself contains multiple steps. The LoosePair strategy also raises questions about fairness and clinical realism, because it relaxes strict OCT-CFP correspondence.

    Overall, the paper is interesting and promising, but the current evidence is not yet strong enough for a higher score.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    Although I believe the paper still has room for further improvement in certain aspects, the authors have indicated in their response that revisions will be made in the final version. I consider that it offers an interesting perspective on optimization for multimodal retinal disease recognition and is worthy of acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper targets multimodal fundus disease recognition, a clinically valuable and technically challenging task. According to the authors, current approaches focus heavily on sophisticated fusion architectures but overlook optimization dynamics during multimodal training. The manuscript claims that optimization conflicts between the OCT and CFP encoders result in inferior performance. To resolve gradient conflicts and enhance cross-modal feature alignment, the authors introduce FGDL and CMSA modules. A loose-pair augmentation strategy is also employed to handle data heterogeneity. The proposed method achieves superior performance over existing methods on two public datasets, MMC-AMD and GAMMA.

  • 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 research topic is well-motivated and clinically significant. Multimodal fundus disease recognition is highly valuable for clinical diagnosis and represents a cutting-edge direction in the medical image analysis community.

    2.The paper identifies a critical and under-explored problem: optimization conflicts in multimodal joint training. It clearly demonstrates that gradient suppression and direction inconsistency between OCT and CFP encoders lead to suboptimal performance. The proposed FGDL and CMSA modules are well-designed and effectively alleviate these issues.

  • 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 literature review is relatively insufficient. Several highly relevant works are not cited or discussed. For instance, the recent MICCAI 2025 paper by Wang et al. on “robust multimodal learning via disentangled representation” is closely related but missing.

    2.The experimental comparisons are not comprehensive enough. The authors only validate their method on a single backbone (ResNet-18). Ablations using different backbones, including representative ophthalmic foundation models such as RETFound, are lacking and should be included for better generalization verification.

    3.The quality of visualization and figures needs improvement. The text in the method diagram is too small and difficult to read under the current paper layout, which reduces readability.

  • 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 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?

    This paper addresses multimodal ophthalmic image analysis, a topic with strong clinical value. It insightfully identifies the optimization conflict issue in multimodal learning and provides novel and targeted solutions. However, the experimental section lacks comprehensive consideration in terms of backbones, and existing foundation models in ophthalmic AI have not been explored as feature extractors. For these reasons, I recommend Weak Accept.

  • 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 strength of this paper lies in its exploration of the promising multi-modal learning problem in ophthalmic images, optimizing the training process from the perspective of gradient conflict. In the rebuttal, the author proposed some solutions to my concerns, such as adding citations of missing literature. Although there are still some non-trivial issues, for example, only ResNet18 is used for comparison, personally, I think it slightly exceeds the borderline for acceptance. Therefore, my decision is to accept this paper.



Review #3

  • Please describe the contribution of the paper

    The authors propose an optimization strategy for multi-model fusion that addresses gradient conflicts between fusion features and unimodal feature optimization. Additionally, they introduce a feature aggregation method, termed cross-model semantic alignment. Combined with a previously proposed loss-pairing strategy for data augmentation, their approach achieves competitive results on two public datasets.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The observation of gradient conflict during multi-model training is insightful, and the proposed method appears to be solid. 2.The proposed method demonstrates significant improvements compared to state-of-the-art approaches.

  • 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 abstract and introduction primarily emphasize FDGL as the main contribution for addressing gradient conflicts. However, the methods section introduces two key components, including CMSA, which appears to be another major contribution. This inconsistency reduces the overall coherence of the paper, and the motivation for introducing CMSA is not sufficiently clear. 2.For the ablation experiments (Table 2), it would be beneficial to include results using FDGL alone, as it is presented as the core contribution of the paper. This would help better isolate and validate its effectiveness. 3.The analysis in Section 3.4 is insightful; however, it remains unclear how resolving gradient conflicts directly translates into improved diagnostic performance. Further clarification or discussion would strengthen the paper. 4.The definition of symbols should be clarified. For example, the meaning of X_m in the algorithm is not clearly defined.

  • 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.

    (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?

    quality of writing, validation of the main contribution

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The authors have addressed some of my concerns, and I acknowledge the novelty of this work. However, the main issue remains that the paper is not yet written to a publishable standard, especially regarding the hierarchy and presentation of the contributions. For example, the authors need to substantially revise the abstract and introduction, which cannot be adequately addressed through a minor revision. Therefore, I recommend rejection at this stage.



Author Feedback

We thank the Area Chair and Reviewers for recognizing our work’s clinical value and competitive performance. We have grouped and addressed the major concerns below. 1.FGDL Novelty & Optimization Dynamics (AC1, R1Q1, R3Q3) Unlike PCGrad which projects conflicting gradients on shared backbone parameters for multi-task learning, FGDL employs a fundamentally decoupled architecture. Following our proposed mathematical logic, the total loss strictly consists of fusion loss, while unimodal loss updates only unimodal backbones. Without FGDL, severe gradient suppression causes unimodal encoders to underfit. Stop-gradient prevents this suppression but isolates the encoders. Our projection step acts as a critical bridge, utilizing the fusion gradient strictly as a directional anchor to rectify isolated unimodal gradients. This decoupling restores learning capacity, enabling encoders to extract highly discriminative features for accurate diagnostic fusion. 2.Evidence of Gradient Conflict (AC3, R1Q5, R1Q6) Rigorous directional evidence of gradient conflict was provided in Fig. 4.As shown in Fig. 4(a), without FGDL, the cosine similarity between unimodal and fusion gradients drops below zero in later epochs, strictly proving the existence of directional conflict. Furthermore, Fig. 4(b) proves our projection step successfully forces this gradient distribution back into the positive (aligned) region. We will improve the text and captions to ensure this critical connection is immediately clear to readers. 3.Module Roles (R1Q3, R3Q1) We agree FGDL is our primary methodological contribution, with CMSA as a supporting module; we will revise the abstract to accurately reflect this hierarchy. 4.Ablations(AC2, R1Q4, R2Q2, R3Q2) On the MMC-AMD dataset, the accuracy rate achieved by using only “baseline + FGDL” was 85.31%, which was an improvement of 9.79% compared to the baseline, demonstrating its effectiveness. And We have verified the generalization ability of this framework at different backbone. However, due to space constraints, this was not included in the paper. Only the most classic baseline was presented. In strict compliance with MICCAI’s policy prohibiting the inclusion of new experimental results during the rebuttal phase, we cannot present the exact numerical metrics here.we firmly commit to including an expanded ablation table in the camera-ready version. This table will rigorously isolate the individual contributions of decoupling, projection, FGDL, CMSA,and LoosePair across multiple backbones (including RETFound). 5.LoosePair Strategy (AC4, R1Q2) LoosePair is primarily employed as a powerful multi-modal data augmentation technique. By allowing category-consistent pairing during training, it expands the combinatorial scale from N to N_cat^2.This deliberately breaks strict instance-level coupling, forcing the model to learn robust, category-invariant pathological representations rather than memorizing instance-specific noise, mitigating overfitting. To ensure rigorous evaluation and clinical consistency, all testing processes are strictly conducted on 1-to-1 paired images (OCT and CFP from the same eye). 6.Literature, Visuals & Symbols (R2Q1, R2Q3, R3Q4) We will certainly include the recent MICCAI 2025 paper by Wang et al. and other highly relevant literature in the revised manuscript. While they resolve multimodal challenges via feature-level disentangled representation, we tackle optimization fundamentally from a gradient perspective; these paradigms are complementary. We will redraw Fig. 2 as a vector graphic with enlarged annotations to clarify the data flow and architectural details.The symbol X_m denotes the intermediate feature map extracted by the backbone, where the subscript min{oct, cfp} specifies the modality. We will carefully examine all the symbols in the text and ensure that there are clear definitions for each one when it first appears.



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 relevant problem in multimodal fundus disease recognition by focusing on optimization conflicts during OCT-CFP joint training, rather than only proposing a fusion architecture. Reviewers found the motivation clinically meaningful and the reported results competitive on two public datasets. However, several concerns remain. The novelty of FGDL is not fully clear with respect to existing gradient projection or stop-gradient methods, and the contributions of FGDL, CMSA, and LoosePair are not sufficiently disentangled. Reviewers also questioned the clinical realism of the LoosePair strategy, the limited evidence for actual gradient conflict, and the lack of experiments with stronger backbones or ophthalmic foundation models. The authors should address:

    1- Novelty of FGDL relative to prior gradient surgery/stop-gradient methods. 2- Clearer ablations isolating FGDL, CMSA, projection, decoupling, and LoosePair. 3- Stronger evidence that gradient conflict exists and is resolved. 4- Clinical validity and fairness of the LoosePair strategy.

  • 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 main pre-rebuttal concerns were about novelty and evidence separation: whether FGDL is sufficiently distinct from prior gradient-surgery / stop-gradient ideas, how much gain comes from FGDL versus CMSA or LoosePair, and whether the evidence for true gradient conflict was rigorous enough. There were also secondary concerns about missing literature, limited backbone diversity, and presentation clarity. The rebuttal clarified the intended hierarchy of contributions, positioned FGDL more clearly relative to prior work, explained the role of LoosePair as training-time augmentation rather than evaluation-time relaxation, and pointed to directional evidence of gradient conflict through cosine-similarity analysis.



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 presents a technically sound contribution to multimodal retinal disease recognition and explores an interesting optimization perspective through gradient conflict mitigation. The rebuttal adequately addresses most of the major concerns raised during review, which is reflected in the updated recommendations from Reviewers 1 and 2. While concerns remain regarding the presentation and organization of the manuscript, I consider these issues addressable through revision and do not view them as fundamental weaknesses of the work itself. Overall, the technical contribution and experimental results are sufficient to support acceptance.

    Recommendation: Accept.



Meta-review #3

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

    Reject

  • Please justify your recommendation.

    This paper addresses an important problem in multimodal fundus disease recognition and presents competitive results on two public datasets. Reviewers appreciated the optimization-oriented perspective and the clinical relevance of the task. However, significant concerns remain regarding the paper’s methodological novelty and the clarity of its contributions. In particular, FGDL appears closely related to existing gradient projection and stop-gradient strategies, while CMSA largely relies on established cross-modal attention mechanisms. The individual contributions of FGDL, CMSA, and LoosePair are not sufficiently disentangled, making it difficult to assess the source of the reported gains. In addition, the presentation and organization of the paper require substantial improvement, particularly regarding the hierarchy of contributions, motivation of CMSA, and overall framing in the abstract and introduction. Although the rebuttal clarified several points, it did not fully resolve these concerns. Therefore, I do not find the novelty and technical contribution sufficiently established for acceptance.



back to top