Abstract

Fundus fluorescein angiography (FFA) is critical for assessing retinal vascular abnormalities, but its acquisition is invasive and not always feasible. In contrast, color fundus photography (CFP) is non-invasive and widely accessible, which has motivated studies on CFP-to-FFA synthesis. However, prior works rely solely on CFP surface texture, fundamentally limiting the ability to reconstruct functional vascular information and subtle pathological changes. To address this, we propose a novel framework that synthesizes FFA from CFP with structural guidance provided by optical coherence tomography (OCT). We construct a multi-modal retinal imaging dataset with paired CFP, FFA, and OCT from 3,676 patient eyes—the first tri-modally aligned dataset in retinal imaging. To bridge the spatial gap between OCT and fundus modalities, we propose a Spatially Aligned Cross-Modal Fusion (\textbf{SACMF}) module that projects depth-resolved OCT features onto the fundus plane and injects them into the CFP encoder via adaptive layer normalization. Beyond feature fusion, we further introduce Token-wise Cross-Modality Alignment (\textbf{TCMA}), a token-level contrastive learning strategy that explicitly aligns CFP and FFA representations at corresponding spatial positions. Our method achieves superior synthesis performance compared to state-of-the-art methods. Moreover, extensive experiments demonstrate that the FFA images synthesized by our approach bring greater improvements in downstream disease diagnosis performance than existing methods, highlighting the clinical potential of our approach as a non-invasive decision-support tool in routine workflows. The code is available at \href{https://anonymous.4open.science/r/OCT-guide-FFA-Syn-D6CE}{here}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1951_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/while-plus/OCT-guide-FFA-Syn

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MaTen_Propagating_MICCAI2026,
        author = { Ma, Tengfei AND Wu, Ruiqi AND Zhang, Chenran AND Geng, Ye AND Su, Na AND Duanmu, Xiangyuan AND Zhou, Tao AND Zhou, Yi AND Fan, Wen},
        title = { { Propagating Structural Guidance: Synthesizing Fluorescein Angiography from Fundus Images and Sparse OCT Scans } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16890},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a multimodal image synthesis framework for generating FFA from color fundus photographs with auxiliary structural guidance from sparse OCT. Its main contribution is to extend CFP-to-FFA synthesis from a single-modality image translation task to a structurally guided multimodal pipeline. Specifically, the paper constructs a tri-modal dataset of 3,676 aligned CFP, FFA, and OCT cases, introduces SACMF to project OCT features onto the fundus plane and propagate structural information through CFP features, and proposes TCMA to align CFP and FFA tokens at corresponding spatial positions.

  • 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 idea of synthesizing FFA from CFP with OCT guidance is clinically meaningful, since both OCT and CFP are non-invasive modalities and therefore more feasible in real-world practice. 2.Incorporating OCT as an auxiliary modality provides additional structural supervision, which plausibly contributes to the improved synthesis performance. 3.The proposed method is designed with clinical considerations in mind, and the downstream experiments provide further evidence of its effectiveness and potential utility.

  • 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 evaluation is limited to a single dataset, making it difficult to assess the robustness and generalizability of the proposed method across different clinical centers, populations, and imaging devices.

  • 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.

    (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?

    good work for clinical use

  • 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 #2

  • Please describe the contribution of the paper

    The authors propose investigating CFP-to-FFA synthesis by incorporating OCT as auxiliary structural guidance, constructing the first paired tri-modal retinal dataset with 3,676 aligned CFP, FFA, and OCT cases. To effectively align and fuse OCT with CFP, an OCT-guided feature modulation module is introduced. Furthermore, the authors implement a token-wise contrastive learning strategy to enforce CFP-FFA alignment. Experimental results demonstrate that the proposed method generates FFA images with higher visual fidelity than existing methods in the literature. Finally, the authors validate that the synthesized FFAs effectively enhance performance in downstream classification tasks.

  • 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.

    Novelty of Dataset: The authors propose a novel method by incorporating OCT into CFP-to-FFA synthesis and construct a large-scale, paired tri-modal retinal dataset, which fills a significant gap in the field. Methodological Soundness: The Trajectory-Aware En-face Projection is effectively designed by accounting for the inherent acquisition differences between OCT and CFP. Comprehensive Experimental Design: The evaluation extends beyond standard visual fidelity. By validating the synthesized FFAs in downstream classification tasks, the authors demonstrate the practical utility of the generated images. Furthermore, the inclusion of a discussion on the ratio of OCT during training provides convincing evidence for the effectiveness of OCT guidance.

  • 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.

    Limited Comparison to State-of-the-Art: While the authors cite relevant works such as “ Lesion-aware generative adversarial networks for color fundus image to fundus fluorescein angiography translation” and “ Non-invasive to invasive: Enhancing FFA synthesis from CFP with a benchmark dataset and a novel network”, a quantitative comparison against these specific baseline methods is absent. Lack of Statistical Rigor: There is no statistical evaluation for the classification performance; the absence of standard deviations makes it difficult to assess the stability and reliability of the reported improvements. Lack of Sensitivity Analysis: The experimental section is missing a parameter sensitivity analysis, particularly regarding the hyper-parameter β in the loss function. It remains unclear how different weights affect the synthesis performance. Insufficient Discussion on Limitations: The qualitative analysis focuses exclusively on successful cases. The paper would benefit from a dedicated section on failure cases and a discussion of the model’s boundary conditions. Lack of Visionary Discussion: There is limited discussion on the distinct clinical roles of CFP, FFA, and OCT. For CFP-to-FFA synthesis to be truly impactful, the authors should discuss whether synthesized FFAs can genuinely substitute for real FFAs in clinical workflows. Incorporating human expert evaluation (e.g., a Turing test by ophthalmologists) would significantly strengthen the argument for its practical application.

  • 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.

    (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?

    To the best of our knowledge, this work is the first to leverage OCT for auxiliary guidance in CFP-to-FFA synthesis. By constructing a dedicated paired dataset and providing a rigorous experimental validation, the authors offer a substantial and well-rounded contribution to retinal imaging research.

  • 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 presents the first multi-modal CFP-to-FFA (color fundus photography to fluorescein angiography) synthesis framework that incorporates OCT as auxiliary structural guidance. Key contributions include: (1) construction of the first tri-modally aligned retinal dataset (3,676 paired CFP, FFA, and OCT cases) using a rigorous vessel-skeleton-based registration pipeline; (2) a Spatially Aligned Cross-Modal Fusion (SACMF) module that projects depth-resolved OCT features onto the 2D fundus plane via trajectory-aware pooling and propagates them via CFP-guided cross-attention; (3) a Token-wise Cross-Modality Alignment (TCMA) module using a symmetric InfoNCE contrastive loss to enforce fine-grained semantic consistency between CFP and FFA tokens. The framework is GAN-based and demonstrates superior FFA synthesis quality and downstream disease classification performance over prior 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.Strong novelty: The work is the first to use OCT as structural guidance for FFA synthesis. This is clinically well-motivated since OCT is non-invasive and available in scenarios where FFA is contraindicated. 2.Dataset contribution: The tri-modal dataset (3,676 cases with registration pipeline) is a genuine and valuable contribution to the community, especially given the scarcity of paired multi-modal retinal datasets. 3.Well-designed SACMF: The trajectory-aware projection is an elegant solution to the OCT–fundus viewpoint discrepancy. Propagating structural cues via CFP-based attention to unscanned regions is practical and effective. 4.Comprehensive evaluation: Both standard image quality metrics (PSNR, SSIM, FID, LPIPS) and downstream diagnostic utility (disease classification) are evaluated. The code is publicly available. 5.Clinically meaningful upper/lower bound analysis provides context for interpreting results.

  • 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.Limited dataset size and split: With only 3,676 cases and a 60/40 train/test split, the dataset is relatively small for a generative model. The choice of 40% for testing is unusual and reduces training data; the rationale should be discussed. 2.GAN instability not discussed: The paper adopts a GAN backbone without discussing training stability, mode coverage, or failure cases. Given the small dataset, GAN training risks are non-trivial. 3.OCT sparsity assumption: The method assumes sparse B-scans are available alongside CFP. This may not always hold in clinical practice. The sensitivity to number and placement of OCT scans is only partially explored (20%-100% training ratio, but not inference-time availability). 4.Qualitative results are limited: Only 4 rows of examples are shown in Figure 2.More diverse failure cases or challenging pathologies would strengthen the qualitative analysis. 5.TCMA supervision detail: It is not fully clear how the FFA tokens are obtained during inference (when real FFA is unavailable), suggesting TCMA may only be active during training.

  • 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 — No ethics concern identified)

  • 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

    This is a well-executed paper with a clinically motivated problem and genuine dataset contributions. The main concerns are the relatively small dataset size, limited discussion of GAN failure modes, and the practical assumption about OCT availability. The paper would benefit from a clearer description of inference-time behavior of TCMA and a discussion of the dataset collection ethics.

  • 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 is technically sound, novel, and well-written. The primary contributions — the OCT-guided synthesis paradigm, the tri-modal dataset, and the SACMF/TCMA modules — are meaningful and well-validated. The score is held back by the relatively small dataset and the 60/40 train/test split that limits generalization claims, as well as an incomplete discussion of inference-time behavior. If the authors can address the dataset split rationale and clarify TCMA at inference time, this would merit a solid Accept.

  • Reviewer confidence

    Very confident (4)

  • [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

1.Dataset Size and Split Constrain Generalization Thank you for your valuable feedback. Compared to large-scale datasets in natural image domains, the cost of data acquisition in medical image generation tasks is relatively high, and thus data scale remains an important challenge in this field. However, it is worth noting that in the specific task of fundus image to FFA synthesis, the data scales used in existing publicly available studies are generally small. For example:

  • “Diabetic retinopathy grading by digital curvelet transform” includes only 59 cases, covering only normal and diabetic retinopathy.
  • “Lesion-aware generative adversarial networks for color fundus image to fundus fluorescein angiography translation” includes 133 cases, with only BRVO.
  • “Non-Invasive to Invasive: Enhancing FFA Synthesis from CFP with a Benchmark Dataset and a Novel Network” includes 600 cases, covering five disease categories. In comparison, our dataset contains 3,676 cases, with 8 disease categories having more than 100 cases each. Under this task setting, our dataset is significantly larger than existing works and covers a richer distribution of lesions.

Regarding the data split ratio, we adopted a 60/40 train/test split rather than the more common 80/20 or 70/30, primarily to enlarge the test set size, thereby enabling a more reliable evaluation of the model’s generalization ability. Therefore, we aim to more thoroughly evaluate the model’s performance on unseen cases, different lesion types, and varying severity levels through a larger test set.

Additionally, it should be noted that compared to the data scales of 59, 133, or 600 cases in existing works, our dataset contains 3,676 cases. Even with a 60/40 split, the training set size remains significantly larger than those used in existing methods.

2.Clarification on TCMA Supervision and Inference Thank you for pointing this out. TCMA is used only during training as an auxiliary supervision module, leveraging ground-truth FFA tokens to provide cross-modal constraints. During inference, the model takes only CFP as input and does not require FFA tokens. We have clarified this distinction in the revised manuscript

3.Insufficient Discussion on Limitations and Failure Cases Thank you for this suggestion. We have added failure case analysis and a discussion of model limitations in the Qualitative Comparison section of the revised manuscript. Meanwhile, we have illustrated a representative failure case in Fig. 2(b) to provide a clear visual example. In this discussion, we explicitly acknowledge certain limitations of our model: (1) vessel boundaries may appear blurred, and (2) the model struggles to recover very small or widely distributed leakage regions, resulting in incomplete synthesis.

4.Undiscussed GAN Training Instability First, with proper architectural design and training techniques, the training stability of GAN frameworks is no longer a major issue. Our GAN model exhibits stable training behavior without evident failure phenomena such as mode collapse. Second, we have also carefully considered and experimented with Diffusion-based methods. As shown in Table 1, we implemented BBDM as a comparative method, but its performance was notably inferior to the GAN-based version. We further attempted to improve its performance by increasing the number of training epochs, yet the gains were not significant. We attribute this primarily to the fact that, in the absence of a large-scale pretrained VAE, directly modeling in pixel-level space poses greater challenges for Diffusion methods, whereas GANs demonstrate better adaptability and stability in such medical image cross-modal generation tasks.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 proposes a clinically meaningful multimodal framework for synthesizing FFA from CFP with structural guidance from OCT. Reviewers agree that the problem is highly relevant, and the approach is novel, particularly through the integration of OCT guidance and the construction of a tri-modal dataset. The method is well-designed and supported by a comprehensive evaluation, including downstream clinical tasks. The concerns are mostly about strengthening the experimental analysis and clarifications rather than fundamental flaws.



back to top