Abstract

Generative image models such as diffusion models can improve performance on clinically relevant tasks by offering deep learning models supplementary training data. However, most conditional diffusion models treat disease stages as independent classes, ignoring the continuous nature of disease progression. This mismatch is problematic in medical imaging because continuous pathological processes are typically only observed through coarse, discrete but ordered labels as in ophthalmology for diabetic retinopathy (DR). We propose an ordinal latent diffusion model for generating color fundus images that explicitly incorporates the ordered structure of DR severity into the generation process. Instead of categorical conditioning, we used a scalar disease representation, enabling a smooth transition between adjacent stages. We evaluated our approach using visual realism metrics and classification-based clinical consistency analysis on the EyePACS dataset. Compared to a standard conditional diffusion model, our model reduced the Fréchet inception distance for four of the five DR stages and increased the quadratic weighted $\kappa$ from 0.79 to 0.87. Furthermore, interpolation experiments showed that the model captured a continuous spectrum of disease progression learned from ordered, coarse class labels.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3874_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/berenslab/OrdinalDiffusionModels

Link to the Dataset(s)

See paper

BibTex

@InProceedings{SchGus_Ordinal_MICCAI2026,
        author = { Schmidt, Gustav AND Berens, Philipp AND Müller, Sarah},
        title = { { Ordinal Diffusion Models for Color Fundus Images } },
        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

    The authors propose to use ordinary labels to represent disease stages in generative modelling, which are inherently progressive. They show that simple equidistant labeling is sufficient to learn such structure.

  • 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 approach is simple and effective. Especially Figure 4 showcases that the ordinal labels are really doing what they are supposed to be doing.

  • 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.
    • There are other diffusion models working with ordinal labels, e. g. Cho et al. “Conditional Diffusion with Ordinal Regression: Longitudinal Data Generation for Neurodegenerative Disease Studies” ICLR 2025 – https://github. com/Hannah37/ConDOR-ICLR25 weakening the claim “our model is the first reproducible ordinal diffusion model to explicitly encode ordinal and structural constraints”

    • I am not sure, why learned margins are included as they provide little benefit even over the baseline and perform consistently worse than equidistant margins. However, apart from this the method section is largely a rehash of existing methods. The section on Ordinal Disease stage conditioning, which is the novel part of the paper is just 3 sentences.

    • Complete absence of any baselines. There should be at least a GAN based one, as StyleGANs have been used for such conditional generation for quite a while now.

    • The images are significantly downscaled and hence unsuitable for downstream training, as the resolution is just 256x256.

    • Furthermore, the vessel structure is not retained in the generated images, e. g. there is additional branching and closing of the vessels. However, this is never commented on.

    • The writing is also not ideal. The first sentence of the abstract starts with “It has been suggested” –> either someone claimed this and then cite the paper or do not mention this.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the proposed method performs as intended, the complete lack of baselines and structural problems with the writing, as well as low methodological novelty lead to my recommendation.

    In the current state, I feel like this would be a good contribution to a workshop.

  • 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 acknowledge that there are multiple ordinal generation works that can easily be adapted to serve as baselines. Nevertheless, the result section misses a comparison to any state-of-the-art method. Not a single model dedicated to the generation of different disease stages was included. Even the rephrased version of the contribution statement is at best missleading as there are already more sophisticated ordinal generation methods, [29] even evaluated on the same dataset.

    The writing is at times poor and a lot of space is dedicated to related work, which is never used as a comparison and further misses a lot of ordinal generation works. The contribution is poorly described as the section about ordinal conditioning is just 3 sentences long. Equidistant Ordinal labels have been used for at least 10 years to generate images from continuous settings, see e.g. InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets by Chen et al.

    The clinical utility is questionable at best, given the hallucinations in vessel structure. It is also unclear if the calculation of the FID score is actuallly showing what it is supposed to show as the reference set contains images from all classes.

    I am voting to reject this paper, as the experimental section is not sufficient given the very limited novelty.



Review #2

  • Please describe the contribution of the paper

    The paper proposes an ordinal latent diffusion model for generating color fundus images conditioned on diabetic retinopathy severity grades. The main contributions are twofold: (1) ordinal disease-grade conditioning, where disease stages are embedded using either equidistant or learned margins rather than standard one-hot encoding, and (2) the combination of this ordinal conditioning with structural conditioning via a contrastive learning-based encoder, enabling independent control over disease severity and retinal anatomy during generation.

    The authors evaluate six model variants (one-hot encoding, 2 novel disease encodings, with or without structural conditioning) on 127,144 fundus images from the EyePACS dataset, measuring FID and QWK of a pre-trained classifier. The equidistant margins model with structural conditioning achieves the best overall performance, reducing FID across four of five DR stages and improving QWK from 0.79 to 0.87 compared to the baseline. Interpolation experiments suggest the model captures a continuous disease spectrum.

  • 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.

    S1.Strong evaluation: The paper presents a thorough experimental design, evaluating six independently trained model variants. S2.Novelty of the combined approach: The combination of ordinal disease-grade conditioning with structure-preserving conditioning is, to the best of this reviewer’s knowledge, novel in the context of diffusion-based medical image generation. The ordinal encodings (equidistant and learned margins) are well-motivated and provide a principled way to exploit the ordered nature of DR severity. S3.Clear improvement over baseline: The equidistant margins model with structural conditioning shows consistent FID and QWK improvements. S4.Reproducibility: The authors provide clean code and model weights. S5.Extendability: The proposed approach is not specific to DR or fundus imaging and could be applied to other medical domains with ordinal severity grading (e.g., cancer staging), which broadens the potential impact of the work. S6.Additional validation experiments: The interpolation and counterfactual generation experiments go beyond standard FID/QWK evaluation and attempt to demonstrate that the model captures a continuous disease spectrum, which strengthens the paper’s narrative (though see C2 and C3 regarding their execution).

  • 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.

    W1.Dataset reproducibility gap: The authors report using 127,144 images from EyePACS after quality filtering, but cite only the original 2009 system description paper. The publicly available versions of this dataset (e.g., https://www.kaggle.com/c/diabetic-retinopathy-detection) contain around 35k labelled samples. For reproducibility, the authors should clarify which version was used, how it can be accessed, and provide the per-class distribution after filtering. W2.FID methodology unclear for minority classes: The authors state that FID was computed between 10,000 real and 10,000 generated images per DR grade. However, the publicly available version has fewer than 1000 samples per stage 3 and 4.Even with the author’s larger 127,144-image version of EyePACS, it seems unlikely that these stages have 10,000 samples each. The authors should clarify how FID was computed for these minority classes. W3.Unfair characterisation of prior work: The paper claims to be the ‘first reproducible ordinal diffusion model to explicitly encode ordinal and structural constraints.’ While the structural conditioning component is novel, the discussion of prior ordinal diffusion work is not entirely fair. Reference [29] is described as not consistently yielding better performance in terms of realism, yet their reported results show improved FID over the standard diffusion baseline on both evaluated datasets. Reference [17]’s Additive Ordinal Embedder enforces ordinal structure through cumulative summation, which is conceptually similar to the learned margins proposed in this paper - dismissing it as having ‘no constraints enforced’ misrepresents that work. The authors are encouraged to more accurately acknowledge prior work. W4.Baseline classifier performance not reported: The QWK is computed using a pretrained DR classifier, and the authors note that this classifier performed better on generated images than on real data. However, the baseline classifier performance on real images is not reported, making it difficult to contextualise the QWK results. This also raises the question of whether the generated images capture the full diversity of real pathology or tend toward more prototypical examples.

  • 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

    C1.Discussion of learned margins: The learned margins model underperforms equidistant margins across nearly all metrics. The paper briefly acknowledges this but does not explore possible explanations. A deeper analysis would strengthen the paper, particularly since learned margins are presented as a key contribution. C2.Interpolation experiment: The interpolation analysis in Figure 4 is an interesting addition, but could be improved. First, the figure is not referenced in the corresponding text. Second, the x-axis label ‘Generated Class’ is misleading - it actually represents continuous conditioning values along the ordinal spectrum, and clearer labelling would help with interpretation. Third, the paper would benefit from explicitly noting that this type of experiment is not possible with the one-hot baseline, since one-hot encoding does not support continuous interpolation - this is itself an advantage of the proposed approach that deserves emphasis. C3.Counterfactual generation lacks quantitative evaluation: The counterfactual experiment presents only a single qualitative example with no numerical evaluation. Reporting FID or QWK for counterfactual generations would have provided stronger evidence. C4.Minor presentation issues: The fire and snowflake emojis in Figure 1 (indicating trainable and frozen components) are not explained in the caption. A more detailed caption describing the model architecture would be appreciated. C5.Suggestion for future work: The authors could strengthen their evaluation of generation diversity by reporting metrics such as IRS or LPIPS pairwise distance, which would help assess whether the generated samples capture the full diversity of real pathology.

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

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

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

    The paper presents a novel combination of ordinal disease-grade conditioning and structural conditioning for diffusion-based fundus image generation, which is well-motivated and, to this reviewer’s knowledge, new in this context. The experimental design is thorough, with six model variants evaluated systematically, and the results demonstrate clear improvements over the baseline. Code and model weights are provided, supporting reproducibility.

    The main concerns relate to the methodological transparency: the dataset version and access route are unclear (W1), and the FID computation for minority classes needs clarification (W2). While this reviewer believes the relative improvements between models likely hold regardless, a paper should clearly describe its data and evaluation protocol to meet the standards expected for publication. Additionally, the related work discussion would benefit from a fairer characterisation of prior ordinal diffusion models (W3), though this is a less critical concern.

    Overall, the paper makes a meaningful contribution to ordinal medical image generation with clear potential for extension to other domains. Subject to a satisfactory rebuttal addressing W1 and W2, this reviewer would be willing to increase the score.

  • 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 satisfactorily addressed my main reproducibility concerns. The authors clarified the dataset version (licensed EyePACS, 2022) and provided the per-class distribution (Grade 0: 100,584; 1: 9,948; 2: 14,823; 3: 960; 4: 829) (W1). They also clarified the FID protocol: a fixed reference pool of 10,000 random real images compared against 10,000 generated per class, rather than per-class real samples (W2). This is a reasonable workaround given the very small Grade 3/4 sample sizes, but the submitted paper’s description (“between 10,000 generated and 10,000 real images for each DR grade”) is misleading and must be corrected in the camera-ready. The relative model comparisons remain valid, and the QWK results are unaffected, so the core claim is supported. The authors narrowed their overclaim to “first reproducible work to explicitly encode ordinal geometric constraints into the conditioning representation of a latent diffusion model for disease staging,” which is defensible and addresses W3.They committed to all requested presentation fixes. The contribution is modest but real and reproducible, with potential to extend to other ordinal grading tasks. Subject to the camera-ready corrections noted, I support acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper proposes an ordinal latent diffusion model for generating diabetic retinopathy (DR) fundus images that explicitly encodes the ordered structure of disease severity into the generation process. Instead of standard one-hot class conditioning, the authors implement two ordinal conditioning strategies: (1) equidistant margins (ci = i) and (2) learned margins (ci = Σv²j, enforcing monotonicity). Additionally, a structural encoder trained via contrastive learning (NT-Xent on a ResNet-50) is used for a dual-conditioning strategy that separates anatomy from pathology, enabling counterfactual image generation. The model is evaluated on EyePACS (127K images) using FID per DR grade and quadratic weighted κ (QWK) on a pretrained DR classifier, showing improved realism and disease consistency over a categorical baseline.

  • 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.Clinically motivated and clearly articulated problem: Modeling disease severity as a continuum rather than discrete categories is a genuine and well-reasoned contribution, particularly relevant for DR where progression is inherently gradual. 2.Simple yet effective design: The ordinal encoding (especially equidistant margins) is minimal and elegant, adding little complexity over baseline while yielding consistent improvements (QWK: 0.79→0.87). 3.Reproducible: Code is publicly available (anonymized link provided), and the full configuration is described. This is a significant plus. 4.Interpolation experiment: The interpolation analysis (Figure 4) is scientifically interesting and provides strong evidence that the model has learned a continuous disease manifold, not just discrete clusters. 5.Honest evaluation: The paper candidly acknowledges that the DR classifier performed better on generated images than real ones, suggesting possible dataset label noise, which is a thoughtful and credible observation.

  • 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 resolution (256×256): Fundus images at this resolution miss fine-grained pathological details (microaneurysms, small hemorrhages). The paper acknowledges this but does not provide experiments at higher resolutions, which is a significant limitation for clinical utility. 2.Learned margins did not consistently outperform equidistant: Despite their theoretical appeal, learned margins performed worse on several metrics. The paper lacks a deeper analysis of why this is the case. 3.No human/expert evaluation: The evaluation is entirely automated (FID + classifier). Clinical utility of generated images requires ophthalmologist assessment, which is explicitly deferred to future work. 4.Structural encoder generalization: The structural encoder is trained on EyePACS (same distribution). Its behavior under domain shift is not evaluated. 5.Proliferative DR generation is poor (FID=38 with equidistant, structural cond.), likely due to class imbalance. The mitigations beyond weighted sampling are not explored.

  • 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 solid, well-scoped methodological paper. It addresses a genuine gap (ordinal modeling in generative models for medical images), provides reproducible code, and offers an interesting interpolation analysis. The main limitations are resolution (256px), lack of expert clinical evaluation, and the underperformance of the learned-margin variant without a satisfying explanation. The paper is a good fit for MICCAI as a focused methodological 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.

    (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 makes a clear and reproducible methodological contribution to ordinal generative modeling for medical images. The equidistant margin model consistently outperforms the categorical baseline and demonstrates meaningful interpolation between disease stages. The score is limited by: (1) the low image resolution which affects clinical relevance, (2) the absence of expert evaluation, and (3) the inconsistent performance of the learned-margin variant without sufficient explanation. The paper is above the acceptance threshold and would benefit from clarification of the learned-margin results.

  • 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.

    The rebuttal addresses the main concerns I and the other reviewers raised, in particular the positioning relative to prior ordinal diffusion work ([29], [17]) and the design choice for FID computation on minority classes. The narrowed novelty claim — explicit scalar geometric ordinal conditioning, fully reproducible — is defensible and clearly delimited from ConDOR’s longitudinal-trajectory setting, [29]’s image-difference regularisation, and [17]’s unconstrained high-dimensional look-up. The dataset clarification (licensed EyePACS, post-quality-filter class counts, Kaggle subset as a reproducible entry point) is sufficient. The decision to assess class-specific fidelity via a pretrained DR classifier rather than per-class FID is methodologically sound given FID’s small-sample bias. The reframing of the baseline question — that this is an ordinal-vs-categorical conditioning study at fixed architecture, not a diffusion-vs-GAN benchmark — is the correct framing for isolating the contribution. The explanation for why learned margins underperform equidistant ones (additional learnable parameters with weak gradient signal from heavily underrepresented late-stage classes) is plausible and consistent with the reported class distribution. The acknowledged limitations (256×256 resolution, absence of expert clinical reading, residual artifacts in proliferative-DR samples) are scope limitations rather than flaws in execution. The QWK gain from 0.79 to 0.87, the FID improvements on four of five grades, the interpolation analysis providing direct evidence of a continuous learned severity manifold, and the availability of code together constitute a useful, methodologically clean contribution. I recommend acceptance.



Author Feedback

Our core contributions - explicit ordinal encoding of disease severity and dual conditioning - were recognized as simple yet effective, novel, and broadly applicable. The interpolation analysis further demonstrated that the model captures a continuous disease spectrum, directly supporting our central design goal. We emphasize that our focus is on ordinal conditioning, where our approach consistently improves the baseline across metrics rather than on generating optimal fundus images.

1.Novelty relative to prior work (R1, R2) ConDOR (Cho et al.) uses ordinal regression to predict disease stage from image features and age, whereas we use the grade label directly as a geometrically structured conditioning representation. ConDOR also targets subject-level trajectories from paired longitudinal data; we address cohort-level generation from cross-sectional data. Both support continuous interpolation, but along different axes: ConDOR along time, ours along disease severity. [29] and [17] impose ordinal structure implicitly - via regularization on image differences and a high-dimensional learned embedding (768-d) - encouraging but not guaranteeing ordinality. We impose it explicitly on a scalar axis, defining an ordinal scale by construction. Notably, [29] does not consistently improve across all DR grades (Fig. 4). We will narrow our claim to: “To our knowledge, this is the first reproducible work to explicitly encode ordinal geometric constraints into the conditioning representation of a latent diffusion model for disease staging.”

2.Dataset (R2) We used the licensed EyePACS dataset (accessed 2022), larger than the public Kaggle subset (~35k images). Since our method requires only DR grade labels, Kaggle offers a reproducible starting point. Class distribution after quality filtering: Grade 0: 100,584; Grade 1: 9,948; Grade 2: 14,823; Grade 3: 960; Grade 4: 829 (N=127,144).

3.Minority class FID computation (R2) FID is highly sensitive to small sample sizes, making class-specific comparisons infeasible. We therefore computed FID using a fixed reference pool of 10,000 randomly sampled real images from the full set and compared against 10,000 generated images from each class. Class-specific feature fidelity was assessed with a DR classifier. We will clarify this in the final version.

4.Stronger baselines (R1) Our core claim is that ordinal conditioning improves over categorical conditioning, not that diffusion models outperform GANs. Our categorical diffusion baseline isolates this by keeping the architecture fixed, ensuring any performance difference is attributable to the architecture-agnostic conditioning alone.

5.Clinical limitations (R1, R3) Ordinal conditioning is resolution-agnostic and directly applicable to higher-resolution LDMs. Expert evaluation and higher resolution are planned next steps.

6.Failure modes (R1, R3) Discussion of vessel structure limitations will be added. Class imbalance is likely the primary driver of higher minority-class FIDs; sampling strategies are a promising future direction.

7.Learned margins (R1-R3) Learned margins make no assumptions about inter-stage distances, allowing the model to learn inter-stage geometry from data, but may suffer from training instability due to the additional learnable parameters and insufficient gradient signal from minority classes. While performing on par with the baseline, they provide improved interpretability through clinically meaningful margins. We will clarify this in the final version.

8.Misc. We will: optimize the abstract (R1); report baseline classifier performance, expand the Fig. 1 caption, and improve Fig. 4 with clearer labels, an explicit text reference, and a note that interpolation analysis is impossible with the categorical baseline (R2). Future directions include quantitative evaluation of counterfactual images via FID/QWK (R2) and latent space analysis for potential bias in the structural encoder (R3).




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 proposes an ordinal latent diffusion model for color fundus image generation, using ordinal disease-stage conditioning and structural conditioning to better reflect the progressive nature of DR. Reviewers appreciated the simplicity of the approach, the reproducibility/code availability, and the interpolation analysis suggesting a continuous disease spectrum.

    However, the reviews are mixed and raise important concerns. The main issues are the limited methodological novelty, incomplete positioning with respect to prior ordinal diffusion models, lack of stronger generative baselines, and unclear dataset/evaluation details, particularly the EyePACS version and FID computation for minority classes. Reviewers also noted limited clinical utility due to low image resolution, lack of expert evaluation, poor generation for severe DR, and insufficient analysis of why learned margins underperform equidistant margins. The authors should address:

    1- Novelty and positioning relative to prior ordinal diffusion work. 2- Dataset version, class distribution, and access details. 3- FID computation for minority DR classes. 4- Absence of stronger generative baselines. 5- Clinical limitations from 256×256 resolution and lack of expert evaluation. 6- Failure modes, especially vessel artifacts and poor severe-DR generation. 7- Why learned margins underperform equidistant margins.

  • 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.

    In the rebuttal, the authors addressed the main concerns reasonably well. In particular, they clarified the EyePACS dataset version and access conditions, explained the class distribution and the protocol used for FID computation in minority classes, and moderated the novelty claims relative to prior ordinal diffusion work. These clarifications resolve some of the main reproducibility and evaluation ambiguities raised in the reviews. The paper still has limitations. The methodological novelty remains moderate, the comparison to stronger generative baselines could be more complete, and the clinical usefulness of the generated images is still constrained by the 256×256 resolution and the lack of expert evaluation. In addition, severe DR generation remains more challenging, and the learned-margin variant is less convincing than the equidistant formulation. That said, the work is technically sound, reproducible, and provides a meaningful contribution in a relevant and emerging direction. Considering the rebuttal and the balance of reviewer opinions, I believe the paper is above the acceptance threshold.



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.

    Many of the major concerns have been addressed by the rebuttal. That said, this work receives diverging ratings, and there remain concerns such as the presentation and over-claims, limitations of the work. The authors are encouraged to carefully follow the reviews and revise the draft accordingly, if it is eventually accepted by this round of MICCAI.



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 authors have addressed most of the concerns raised by reviewers. I recommend Accept.



back to top