List of Papers Browse by Subject Areas Author List
Abstract
Diabetic retinopathy (DR) progresses as a continuous and irreversible deterioration of the retina, following a well-defined clinical trajectory from mild to severe stages. However, most existing ordinal regression approaches model DR severity as a set of static, symmetric ranks, capturing relative order while ignoring the inherent unidirectional nature of disease progression. As a result, the learned feature representations may violate biological plausibility, allowing implausible proximity between non-consecutive stages or even reverse transitions. To bridge this gap, we propose Directed Ordinal Diffusion Regularization (D-ODR), which explicitly models the feature space as a directed flow by constructing a progression-constrained directed graph that encourages forward disease evolution. By performing multi-scale diffusion on this directed structure, D-ODR imposes penalties on score inversions along valid progression paths, thereby effectively preventing the model from learning biologically inconsistent reverse transitions. This mechanism aligns the feature representation with the natural trajectory of DR worsening. Extensive experiments demonstrate that D-ODR yields superior grading performance compared to state-of-the-art ordinal regression and DR-specific grading methods, offering a more clinically reliable assessment of disease severity. Our code is available at https://github.com/HovChen/D-ODR.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1292_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HovChen/D-ODR
Link to the Dataset(s)
https://github.com/chehx/DGDR/tree/main/GDRBench
BibTex
@InProceedings{CheHua_Directed_MICCAI2026,
author = { Chen, Huangwei AND Jia, Junhao AND Li, Ruocheng AND Yang, Cunyuan AND Li, Wu AND Pang, Xiaotao AND Chen, Yifei AND Wang, Haishuai AND Bu, Jiajun AND Wu, Lei},
title = { { Directed Ordinal Diffusion Regularization for Progression-Aware Diabetic Retinopathy Grading } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes an ordinal regression framework motivated by the idea that ordinal label structure should not necessarily be modeled as a fixed linear or geometrically uniform progression. Instead of assuming that adjacent labels are equally related in representation space, the method learns graph relations dynamically during training from mini-batch samples. The approach is typically built on top of a feature space extracted by strong visual encoders, such as Vision Transformers, and introduces an additional loss term intended to penalize trajectories that violate clinically meaningful ordering constraints, such as implausible reversals in diabetic retinopathy progression. The main contribution is therefore twofold. First, the paper argues for a more flexible representation of ordinal structure through learned sample relations rather than fixed adjacency assumptions. Second, it incorporates a progression-aware constraint in the learning objective, with the goal of respecting clinically plausible ordering. Conceptually, this is an interesting idea. It addresses a real limitation of many ordinal prediction approaches, namely that ordinality is often encoded too rigidly. The mathematical formulation appears correct and the problem is reasonably motivated. However, the core issue is that the experiments do not convincingly demonstrate that this more sophisticated ordinal modeling leads to robust performance gains or to a meaningfully different learned structure than simpler 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 challenges a common simplifying assumption in ordinal regression, namely that ordinal adjacency should be modeled as a fixed linear relation. This is a reasonable and potentially important point, especially in medical severity grading tasks where neighboring labels may not correspond to uniform semantic distances. 2.Based on the manuscript description, the formulation appears technically correct. The graph-learning mechanism and the progression-aware penalty are coherent and seem internally consistent. 3.The loss component penalizing implausible reversals in disease progression is well motivated. This gives the method a task-aware inductive bias rather than adding complexity without justification. 4.The basic idea of learning relational ordinal structure rather than imposing it a priori could be useful in other medical imaging problems involving ordered labels.
- 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 experimental evidence is not convincing enough to support the main claims. This is the central weakness. The paper introduces a relatively elaborate mechanism for graph-based ordinal modeling, but the reported quantitative gains do not appear strong or robust enough to justify the added complexity. 2.In particular, on Messidor, removing the graph component appears to produce only marginal changes relative to the baseline. Based on the reported variability, the standard deviation seems large enough that the average gains are not clearly meaningful. This raises a direct concern about whether the graph-learning block is actually responsible for the reported improvements. 3.Figure 3B is especially unconvincing. If the distribution of learned node degrees is not substantially different from the baseline, then the paper does not clearly show that the learned graph structure is meaningfully different or more informative than simpler alternatives. 4.The limited sensitivity to hyperparameters could be interpreted as robustness, but here it also suggests that the newly introduced coefficients may not be driving behavior in a substantial or identifiable way. 5.The current presentation does not make it clear whether differences between the proposed method and baselines are statistically significant. 6.The link to the repository is not working.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
My recommendation is driven primarily by the mismatch between the paper’s conceptual ambition and the strength of its empirical support. I think the idea is interesting and the mathematical formulation appears sound. The motivation for relaxing rigid ordinal structure and penalizing implausible disease progression is reasonable and potentially useful. However, for a paper centered on a specific methodological contribution, the burden of proof is to show clearly that the contribution matters in practice. The quantitative improvements appear modest, the ablations suggest that the graph component may not be doing very much, and the qualitative evidence does not clearly show that the learned structure differs meaningfully from the baseline. In particular, the fact that the standard deviation appears to absorb much of the mean improvement substantially weakens the central claim.
So my decision is not based on the idea being wrong, but on the evidence being insufficiently convincing for acceptance. In its current form, I see this more as an interesting but preliminary methodological direction than as a strong MICCAI paper.
- 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.
I thank the authors for the detailed rebuttal. I appreciate the clarification that Fig. 3(b) shows predicted severity-score distributions rather than learned node-degree distributions, and I acknowledge that the anonymized code link is available and accessible. I also agree that the limited sensitivity to hyperparameters can be interpreted as robustness rather than irrelevance.
That said, my main concern remains essentially unchanged. The core issue for me is not whether the method induces a somewhat more ordinally structured representation, but whether that added structure translates into a clearly demonstrated practical benefit. In the current paper, the empirical gains still appear relatively modest, and the additional statistical evidence based on three seeds does not fully resolve my concern that the overall improvement is limited in magnitude. The representation-level analysis is supportive, but ultimately indirect: a cleaner ordinal embedding is only compelling if it yields a strong and robust advantage in the main task metrics.
Therefore, while I appreciate the authors’ clarifications and agree that some of my earlier concerns should be softened, I am still not convinced that the paper provides sufficiently strong empirical evidence for acceptance. My overall assessment remains on the reject side.
Review #2
- Please describe the contribution of the paper
This paper proposes Directed Ordinal Diffusion Regularization (D-ODR), a novel regularization framework for progression-aware ordinal regression in diabetic retinopathy (DR) grading. The authors identify that existing ordinal regression methods treat severity levels as static, symmetric ranks, ignoring the inherent unidirectional and irreversible nature of disease progression. D-ODR addresses this by Directed Graph Construction, Multi-Scale Diffusion Regularization, Zero-Inference-Overhead Design.
- 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.
Multi-Scale Diffusion Captures Transitive Relationships, The multi-step random walk formulation, elegantly extends local ordinal constraints to global consistency. By penalizing score inversions along t-step paths, the method enforces transitive monotonicity.
Zero Inference Overhead Design: The regularization module is training-only and discarded at inference, This is a practically important design choice that makes the method deployable in real-world screening scenarios where computational resources may be limited
- 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.
Misalignment Between Title/Claims and Actual Method: Not True “Progression Modeling”, The paper’s title and core contribution claim center on “Progression-Aware” and “Progression Modeling,” but the proposed method does not actually model disease progression in the clinical or temporal sense. This represents a significant discrepancy between the framing and the technical substance. The term “progression modeling” in medical AI literature typically refers to Temporal dynamics, Transition probabilities, Longitudinal trajectory analysis, etc. D-ODR is essentially “Similarity-Weighted Pairwise Ordinal Regression”—it applies stronger penalties to mis-ranked pairs that are close in feature space. The “directed” aspect merely removes backward edges from the pairwise constraints. The actual contribution—adaptive weighting of pairwise ordinal constraints based on feature similarity—is still valid and useful, but it is fundamentally different from progression modeling.
Computational Complexity and Scalability Concerns: The complexity of graph construction and the multi-step matrix power operations are not adequately discussed. For large-scale screening programs with millions of images, this approach becomes computationally prohibitive compared to pointwise methods like CORN.
The paper reports mean±std from only 3 random seed runs. While this provides some indication of variance, (1) n=3 is insufficient for reliable variance estimation, (2) no confidence intervals are provided, and (3) no statistical significance tests (e.g., paired t-test) are performed to validate claims of ‘consistent improvements’ over baselines.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The directed graph formulation with multi-scale diffusion (Equations 2-5) is a norval contribution to ordinal regression. The zero-inference-overhead approach (training-only regularization) makes the method deployable in real-world screening scenarios, addressing a common barrier to adopting sophisticated regularization techniques.
Major factors limiting the score: Critical framing issue: The title and abstract claim “progression modeling” and “capturing transition dynamics,” but the method does not model temporal progression or patient-level trajectories. It operates on cross-sectional data using feature-space similarity (kNN), not temporal progression patterns.
Insufficient statistical rigor: The results report mean±std from only 3 random seed runs, which is insufficient for reliable variance estimation.
- 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
Review #3
- Please describe the contribution of the paper
The paper proposes a progression-aware ordinal learning framework called Directed Ordinal Diffusion Regularization (D-ODR) for diabetic retinopathy (DR) grading. The main idea is to incorporate the clinically proven assumption that DR severity typically progresses in one direction by constructing a directed graph in the learned feature space. In the proposed approach, the edges are allowed only towards samples with equal or higher severity levels. Multi-step diffusion is then applied to propagate ordinal relationships, and a regularization loss penalizes violations of monotonicity along reachable forward paths. The proposed work is evaluated on multiple public datasets using different backbones. The reported results demonstrate consistent improvements over several existing ordinal regression approaches and DR grading 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 is grounded in the clinical observation that DR generally progresses in a largely irreversible manner. Encoding this prior into the learning framework is intuitive and meaningful. 2.The proposed framework is well-structured and relatively easy to implement. 3.The idea of enforcing consistency only along forward severity relations is reasonable and potentially useful for reducing clinically implausible ordering errors. 4.The proposed regularizer is simple and well-integrated with the training phase. It is easy to understand and does not require architectural changes at inference time, which improves practical usability. 5.The contribution of each component is clearly supported through ablation experiments.
- 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.While the introduction of directional constraint is meaningful, the technical novelty appears moderate. The proposed work builds on existing components such as graph-based regularization and diffusion processes. The overall contribution appears incremental rather than fundamentally novel. 2.The paper repeatedly emphasizes on biological plausibility and progression-aware learning. However, these claims are not directly validated through any clinical study or longitudinal data. The method is trained on cross-sectional severity labels rather than true longitudinal progression data. Therefore, the claimed link to disease progression remains indirect and the method is better supported as direction-aware ordinal regularizer than a true progression model. The authors should either soften these claims or evaluate the proposed approach on some longitudinal data for completeness. 3.The proposed method is evaluated only on DR datasets. As a result, it remains unclear whether the approach is generalizable to other ordinal medical tasks. 4.There is no discussion of failure cases. Including qualitative or quantitative analysis of failure cases such as how the model behaves on borderline or adjacent severity cases, would strengthen the work. 5.The experimental section would also be stronger with a clear discussion of training overhead, robustness to label noise and comparison to simpler directional ranking baselines without diffusion to showcase the actual gains coming from the proposed diffusion formulation.
- 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
1.Can the authors clarify how they distinguish between direction-aware ordinal supervision and true disease progression modeling, given that the method uses cross-sectional labels rather than longitudinal patient trajectories? 2.How are the claims of biological plausibility and progression-aware learning justified in the absence of clinical validation or longitudinal data? 3.What is the additional training-time cost in terms of memory and runtime introduced by the graph construction and multi-step diffusion? 4.How robust is the method to label ambiguity or noise, especially between adjacent DR grades? 5.Would the method generalize to other ordinal medical imaging tasks beyond diabetic retinopathy?
- 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?
The paper has a clear clinical motivation, but there are a few limitations in terms of novelty and validation. The contribution appears incremental rather than fundamentally new, as it combines existing components in a task-specific manner. While the progression-aware perspective is relevant and potentially useful, the claims about disease progression are only indirectly supported by the current experiments.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 author response is not able to justify all the comments. After addressing all the comments from the reviewers and incorporating them into the paper, the manuscript will be ready for the publication.
Author Feedback
We thank all reviewers for their constructive comments and for recognizing the motivation, clarity, and practical zero-inference-overhead design of D-ODR. [R1Q1&2&5, R2Q3: empirical evidence]Following previous works [1,2], we report mean±std over 3 random seeds. To address the significance concern without adding new experiments, we conducted paired t-tests on the existing results, which further support that D-ODR consistently outperforms its baselines, with several key comparisons reaching p<0.05 (all QWK comparisons). Regarding R1’s concern on Messidor, D-ODR improves both ViT-B/16 and RETFound, especially in QWK. In Table 2, the full model also shows better performance than ablations, indicating that the graph component is effective. [R1Q3: clarification of Fig. 3(b) and graph-structure] Fig. 3(b) shows predicted severity-score distributions rather than learned node-degree distributions. It visualizes improved ordinal separation of predictions, while Fig. 3(a) further shows that D-ODR produces more ordinally organized embeddings across severity stages. [R1Q4: hyperparameter sensitivity] Fig. 2 indicates robustness to λ, bs and k, rather than irrelevance. Table 2 indicates that removing Dir./FC/MSD changes performance and stability, showing that the regularizer structure is the key factor. [R1Q6: reproducibility] The anonymized code link is provided in the Abstract, and we have rechecked that it is accessible. [R2Q1, R3Q2: framing of “progression-aware”] We agree that D-ODR does not model patient-specific longitudinal trajectories. In this work, “progression-aware” refers to the clinically defined severity-stage ordering of DR rather than longitudinal prediction. Established DR grading systems define DR as ordered stages of disease severity, and stepwise worsening along these scales is widely used in clinical studies to describe DR “progression”[3,4]. Therefore, although our data are cross-sectional, the labels encode a clinically macroscopic disease-stage hierarchy. D-ODR leverages this hierarchy by enforcing direction-aware ordinal constraints and avoiding clinically implausible reverse ordinal relations. [R2Q2, R3Q5: algorithm complexity] The extra memory cost only comes from batch-level transition matrices and scales as O(B^2), without constructing or storing any full-dataset graph. The additional computation is restricted to mini-batch graph construction and diffusion, which is minor compared with the backbone forward/backward pass. In our training runs, the cost remains almost the same as the original backbone. [R3Q1: novelty] We respectfully clarify that the novelty of D-ODR lies in introducing a disease-aware directed ordinal inductive bias that guides diffusion along clinically plausible progression paths. Unlike generic graph smoothing, D-ODR removes biologically invalid reverse relations, yielding a progression-constrained regularizer. [R3Q3: generalizability] The formulation is generic to ordered labels, but current evidence is limited to DR. We are willing to explore other ordinal tasks in future work. [R3Q4: failure case] While a dedicated failure case section is not included in the main manuscript, our qualitative analysis offers insights into typical failure modes. Most errors occur between borderline adjacent grades, where score distributions still overlap. Even in these challenging cases, D-ODR reduces intra-class variance and produces a more monotonic severity ordering.
[1] Lin T et al. (2025). Efficiency and safety of automated label cleaning on multimodal retinal images. npj Digital Medicine. [2] Xu Q et al. (2026). Vision-Language Models Guided Graph Concept Reasoning for Interpretable Diabetic Retinopathy Diagnosis. AAAI. [3] Wilkinson, C. P. et al. (2003). Proposed international clinical diabetic retinopathy and diabetic macular edema disease severity scales. Ophthalmology. [4] Prashant Nair et al. (2016). Report From the NEI/FDA Diabetic Retinopathy Clinical Trial Design and Endpoints Workshop. IOVS.
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 a progression-aware ordinal regularization framework for DR grading, using a directed graph and diffusion-based constraint to enforce clinically plausible ordering. Reviewers agree that the motivation is relevant and the formulation is coherent.
However, important concerns remain. The reported gains appear modest and not clearly statistically convincing, and the ablations do not fully isolate the contribution of the graph/diffusion component. Reviewers also noted a mismatch between the “progression-aware” framing and the use of cross-sectional data rather than longitudinal trajectories.
The authors should address the following points:
1- Provide stronger statistical evidence that the improvements are meaningful and robust. 2- Distinguish more clearly between direction-aware ordinal regularization and true longitudinal progression modeling. 3- Provide evidence that the learned graph captures meaningful ordinal structure. 4- Discuss computational cost, scalability, and practical feasibility. 5- Clarify the evaluation protocol, including variance reporting, number of runs, statistical testing, and the broken repository link.
- 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.
the main concern remains unresolved: the empirical benefit is still not strong enough relative to the added complexity and framing. Even after rebuttal, the gains appear modest overall, and the evidence that the proposed graph/diffusion mechanism yields a practically meaningful downstream advantage remains limited. The rebuttal improves clarity, but it does not fundamentally change the balance of evidence regarding impact. In addition, the “progression-aware” framing still remains somewhat stronger than what is directly supported by cross-sectional label supervision.
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.
The rebuttal successfully resolved a number of misunderstandings in the original reviews. Despite this, reviewers decided not to modify their recommendations, so the majority vote would still be against acceptance. Unfortunately, R2 did not leave any post rebuttal comments, and those of R3 are not very specific. R1 cites “relatively modest empirical gains” as the main reason for sticking to their original recommendation. Considering that reviewers found the idea novel, sound, and interesting, and that Table 1 shows advantages over all four datasets and two backbones, sometimes by more than 10 points, rejecting the paper for this reason would seem overly harsh to me.