List of Papers Browse by Subject Areas Author List
Abstract
Automatic diabetic retinopathy (DR) grading is essential for scalable screening and timely referral to prevent vision loss. However, most existing methods treat DR grading as a multi-class classification problem, overlooking the ordinal nature of disease severity and the asymmetric clinical risk, where underestimation is far more detrimental than overestimation. In this paper, we propose Risk-Averse Ordinal Reinforcement Learning (RAO-RL), a novel framework that reformulates DR grading as a risk-aware decision-making task, enabling the direct optimization of non-differentiable clinical risks. First, we introduce ordinal-inductive feature learning with a strict monotonicity constraint to embed the progressive manifold of DR severity directly into the latent space. Then, we employ the cumulative probability regularization that enforces probabilistic consistency across cumulative severity during pretraining. Finally, we propose a reinforcement learning strategy, called group relative risk-averse policy optimization, to improve accuracy while imposing asymmetric penalties on high-risk misclassifications by clinic-guided rewards. Extensive experiments on four public benchmarks demonstrate that RAO-RL achieves state-of-the-art accuracy while significantly reducing clinically severe errors, offering a safer solution for automatic DR grading. The code is available at github.com/ZhangYH0502/RAO-RL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0916_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaYuh_Adapting_MICCAI2026,
author = { Zhang, Yuhan AND Wang, Xi AND Li, Mingchao AND Ma, Xiao AND Xue, Wufeng AND Ni, Dong},
title = { { Adapting Ordinal-Risk Alignment for Clinically Safe Diabetic Retinopathy Grading } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses DR grading from a risk-aware and ordinal perspective, arguing that standard multi-class classification does not adequately capture the progressive nature of DR severity or the asymmetric clinical cost of underestimating severe cases. To this end, the paper proposes a unified framework, RAO-RL, which combines: (i) an ordinal-inductive feature learning module (OIFL) to inject ordered severity priors into feature aggregation, (ii) a cumulative probability regularization (CPR) pretraining strategy to incorporate ordinal supervision beyond standard cross-entropy, and (iii) a group relative risk-averse policy optimization module (GR²PO) that optimizes a clinically guided reward with ordinal-distance and asymmetric safety penalties. Experiments on four public DR grading datasets show improvements in QWK, F1, and accuracy over recent baselines, while additional analyses on APTOS and DDR suggest reduced underestimation and fewer clinically undesirable misclassifications.
- 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 explicitly considers ordinal structure and asymmetric clinical risk, which is a meaningful perspective for DR grading. 2.The method is organized in a logically consistent way: ordinal-aware feature learning, ordinal pretraining, and a risk-aware optimization stage. The safety-oriented evaluation and reward ablation are also aligned with the paper’s central claims. 3.The paper reports results on four public DR datasets and consistently improves over several recent methods on QWK, F1, and ACC, which suggests the method is competitive in standard benchmark settings. 4.Compared with many papers that only report generic classification metrics, the paper includes error-structure analysis, underestimation metrics, component ablation, and reward ablation, which is a positive aspect of the evaluation.
- 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 introduction emphasizes catastrophic cross-grade errors (e. g. , predicting grade 4 as grade 0) as a central limitation of existing formulations. However, in ordinal grading tasks, the dominant confusion pattern is often between adjacent grades rather than across distant grades. It is therefore unclear whether such extreme errors are truly a representative failure mode, or instead low-frequency but high-cost outliers. This distinction matters because it affects whether the proposed framework is addressing a dominant weakness of existing DR models or a rarer safety-critical failure case. 2.Even if severe cross-grade errors do occur, they may arise from factors such as image quality, insufficient lesion capture, class imbalance, scarcity of severe cases, or limited backbone/representation capacity. If the model struggles to distinguish visually distant stages, the bottleneck may lie more in representation learning than in the decision objective. The paper would benefit from a clearer argument for why the main intervention point should be the training objective and RL optimization. 3.OIFL imposes monotonicity on the scalar biases 𝑏𝑘, but this does not automatically imply that the resulting level embeddings 𝑒𝑘 themselves carry an interpretable progression of DR severity, since 𝑒𝑘 also depends on patch features and their level-wise weights. As currently written, it is not entirely clear whether OIFL is intended mainly as an ordinally biased aggregation mechanism, or whether the authors also claim that 𝑒𝑘 themselves represent severity-specific semantic levels. If the latter is intended, stronger supporting analysis would be helpful. 4.GR²PO is framed as a policy optimization method, but the underlying task remains a single-step static classification problem, and the reward is directly computable from the predicted class and the ground truth. In this setting, it is unclear why a reinforcement learning policy optimization framework must be adopted instead of directly optimizing a supervised objective function with the same structure. In my view, this is one of the core unresolved issues of this paper. 5.This paper demonstrates that AUR/WUR can reduce underestimation and further misclassification, supporting the intended motivation. However, since the reward mechanism explicitly inhibits underestimation of severe cases, there is a risk of over-conservatism—systematically predicting higher grades to avoid dangerous missed diagnoses. Current experiments lack indicators such as sensitivity and specificity to assess whether this method can improve safety without increasing overestimation, reducing specificity, or increasing unnecessary referral burden. 6.Table 2 provides useful reward ablation, but a crucial comparison is lacking: optimizing the same model using a supervised ordinal-aware objective function without reinforcement learning. Without such a comparison, while this paper demonstrates the effectiveness of the approach, it does not fully prove that reinforcement learning itself is necessary. 7.The authors mention that hyperparameters will be provided in the open-source code, demonstrating a strong willingness to open-source. However, for the current version being reviewed, I still recommend reporting at least the most critical hyperparameter settings that significantly affect the method’s behavior in the main text or supplementary materials. This is because a significant portion of the parameters in this paper are not ordinary training details, but rather method-defining parameters that appear directly in the core formulas and key designs, especially those related to reward construction and policy optimization in OIFL, CPR, and GR²PO. If the selection methods for these key parameters are missing in the current manuscript, readers and reviewers will find it difficult to judge the reproducibility of experimental results, parameter sensitivity, and the robustness of the overall conclusions.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 explores an important problem and proposes a coherent risk-aware DR (Digital Ranking Error) grading framework, achieving encouraging improvements on multiple datasets and providing ablation studies. However, I have reservations about several key points: the motivation behind the improvements may overstate the prevalence of catastrophic cross-grading errors; the description of OIFL is not entirely clear; and the paper does not convincingly demonstrate why reinforcement learning is needed in single-step static classification scenarios, rather than a supervised ordinal-aware loss optimization scheme. For these reasons, despite the paper’s interest and potential impact, I believe its current methodological justification is insufficient.
- 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 #2
- Please describe the contribution of the paper
The paper introduces Risk-Averse Ordinal Reinforcement Learning (RAO-RL), a novel framework for Diabetic Retinopathy (DR) grading. The core contribution is the reformulation of DR grading from a standard classification task into a risk-aware decision-making task. Key technical components include:
1.An ordinal-inductive feature learning module with strict monotonicity constraints to capture the progressive nature of DR.
2.Cumulative probability regularization to ensure probabilistic consistency.
3.A reinforcement learning strategy (Group Relative Risk-Averse Policy Optimization) that utilizes clinic-guided rewards to penalize “dangerous” misclassifications (underestimation of severity) more heavily than overestimation.
- 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.Clinical Relevance: The paper addresses a critical gap in automated DR screening, the asymmetric cost of errors. By prioritizing the reduction of underestimation, the work aligns closely with clinical safety requirements.
2.Technical Novelty: The integration of ordinal manifold learning with Reinforcement Learning is a clever and original formulation. Moving beyond traditional cross-entropy loss allows the model to optimize for non-differentiable clinical risk metrics directly.
3.Methodological Rigor: The use of monotonicity constraints and cumulative probability regularization provides a solid theoretical foundation for handling the ordinal nature of disease stages, which is often neglected in standard CNN/Transformer classifiers.
4.Strong Evaluation: The framework is validated across four public benchmarks, demonstrating not just state-of-the-art accuracy, but a significant reduction in clinically severe errors (high-risk misclassifications).
Overall the paper is well written with good clarity and flow of the narrative. The results look promising and solution is clever formulation. The figures and tables are really informative and easy to understand.
- 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.Computational Complexity: While the performance is strong, the three-stage approach (Ordinal learning, Cumulative regularization, and RL optimization) may involve higher training complexity compared to end-to-end classification. A brief discussion on training time or resource requirements would be beneficial.
2.Reward Design Sensitivity: The effectiveness of the RL component likely depends on the specific “clinic-guided rewards.” While the results are impressive, more analysis on how sensitive the model is to different reward weighting schemes would strengthen the paper.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The recommendation is based on the paper’s high clinical impact and methodological innovation. Unlike many DR grading papers that focus solely on incremental accuracy gains, this work tackles the “safety” aspect of AI deployment by penalizing high-risk misclassifications. The formulation is clever, the evaluation is extensive across multiple datasets, and the presentation is top-tier. The minor concerns regarding complexity do not outweigh the significant contribution to making DR grading safer and more trustworthy.
- 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 main contribution of the paper is the proposed Risk-Averse Ordinal Reinforcement Learning (RAO-RL) framework, which integrates ordinal feature learning with risk-aware reinforcement learning to enhance clinical safety in diabetic retinopathy (DR) grading. This method significantly reduces the underestimation of severe DR cases, ensuring higher sensitivity, particularly in patients requiring urgent treatment, while also minimizing extreme misclassifications, improving both the accuracy and reliability of DR grading.
- 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 proposed framework, “Risk-Averse Ordinal Reinforcement Learning (RAO-RL)”, effectively combines ordinal classification with risk-aware reinforcement learning (RL), addressing significant challenges in diabetic retinopathy (DR) grading, such as the underestimation of severe cases. （2） The paper emphasizes clinical safety, reducing the risk of underestimating severe diabetic retinopathy cases, which is crucial for real-world medical applications. （3）Extensive experiments on four public datasets (APTOS, DDR, Messidor-2, EyePACS) demonstrate that the proposed method outperforms existing methods in terms of accuracy, F1 score, and quadratic weighted kappa (QWK). （4）The ablation studies (Table 2) effectively show the contribution of each component, such as Ordinal-Inductive Feature Learning (OIFL), Cumulative Probability Regularization (CPR), and Group Relative Risk-Averse Policy Optimization (GR2PO), clearly demonstrating how each step enhances the model’s performance.
- 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） Although the GR2PO(Group Relative Risk-Averse Policy Optimization) method is introduced in detail, it would be helpful to provide a clearer explanation of how it integrates with the overall model architecture and its impact on performance metrics (not just clinical safety), as this would aid readers who are less familiar with reinforcement learning techniques. （2）While the paper discusses clinical safety, including underestimation rates and misclassification patterns, further explanation of how these results affect real-world medical decisions or expert insights would enhance the clinical relevance of the paper.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper presents an innovative approach to DR grading with a focus on clinical safety. However, more clarification on the practical impact of the reinforcement learning method and its clinical interpretation would improve the paper.
- 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 RAO-RL framework that addresses critical issues in diabetic retinopathy grading. The experimental results demonstrate its superiority over existing methods in terms of accuracy and clinical safety. While some areas, such as the reinforcement learning details, could benefit from further clarification, the overall quality of the paper is high, and it meets the acceptance threshold.
- 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
N/A
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 addresses an important and clinically meaningful problem in diabetic retinopathy grading by explicitly modeling ordinal structure and asymmetric clinical risk. All reviewers agree on the relevance of the problem and the strength of the empirical evaluation, with consistent improvements reported across multiple datasets and metrics, including clinically motivated analyses such as underestimation rates.
The main point of discussion concerns the methodological justification of the reinforcement learning component. Reviewer #1 raises valid questions about whether RL is necessary in a single-step classification setting and whether similar objectives could be achieved with a supervised ordinal-aware loss. Additionally, clarification is needed regarding the interpretation of the ordinal feature representations and the prevalence of catastrophic cross-grade errors. Reviewers also suggest providing more details on reward design, sensitivity, and key hyperparameters to improve reproducibility.
Despite these concerns, the consensus is that the paper presents a well-motivated, technically coherent, and empirically strong contribution with clear clinical relevance. The identified issues are primarily about clarification and strengthening the justification rather than fundamental flaws.