List of Papers Browse by Subject Areas Author List
Abstract
As a long-term complication of diabetes, diabetic retinopathy (DR) progresses slowly, potentially taking years to threaten vision. An accurate and robust evaluation of its severity is vital to ensure prompt management and care. Ordinal regression leverages the underlying inherent order between categories to achieve superior performance beyond traditional classification. However, there exist challenges leading to lower DR classification performance: 1) The uneven distribution of DR severity levels, characterized by a long-tailed pattern, adds complexity to the grading process. 2)The ambiguity in defining category boundaries introduces additional challenges, making the classification process more complex and prone to inconsistencies. This work proposes a novel autoregressive ordinal regression method called AOR-DR to address the above challenges by leveraging the clinical knowledge of inherent ordinal information in DR grading dataset settings. Specifically, we decompose the DR grading task into a series of ordered steps by fusing the prediction of the previous steps with extracted image features as conditions for the current prediction step. Additionally, we exploit the diffusion process to facilitate conditional probability modeling, enabling the direct use of continuous global image features for autoregression without relearning contextual information from patch-level features. This ensures the effectiveness of the autoregressive process and leverages the capabilities of pre-trained large-scale foundation models. Extensive experiments were conducted on four large-scale publicly available color fundus datasets, demonstrating our model’s effectiveness and superior performance over six recent state-of-the-art ordinal regression methods. The implementation code is available at \url{https://github.com/Qinkaiyu/AOR-DR}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0874_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Qinkaiyu/AOR-DR
Link to the Dataset(s)
https://github.com/chehx/DGDR/blob/main/GDRBench/README.md
BibTex
@InProceedings{YuQin_Parameterized_MICCAI2025,
author = { Yu, Qinkai and Zhou, Wei and Liu, Hantao and Xu, Yanyu and Wang, Meng and Zhao, Yitian and Fu, Huazhu and Ye, Xujiong and Zheng, Yalin and Meng, Yanda},
title = { { Parameterized Diffusion Optimization enabled Autoregressive Ordinal Regression for Diabetic Retinopathy Grading } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15974},
month = {September},
page = {455 -- 465}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a novel approach to diabetic retinopathy grading by formulating it as an autoregressive ordinal regression task. Rather than treating DR levels as categorical labels, the authors model the progression as a sequence of binary tokens reflecting ordinal structure, using an autoregressive framework. The proposed method is implemented using parameterized diffusion optimization and integrates both CNN and ViT-based image feature extractors. Code is publicly available, which enhances reproducibility.
- 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 core idea of modeling DR grading as an autoregressive sequence is original and conceptually sound.
The method provides a fresh perspective on ordinal prediction tasks in medical imaging.
The paper is clearly written, with a coherent structure and good technical detail.
Public release of code adds to the transparency and reproducibility of the work.
Results are well presented, and the approach is competitive with state-of-the-art baselines.
- 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 model generates ordinal sequences (e.g., [1,1,0,0]), but there is no explicit mechanism to prevent invalid sequences like [0,1,0,1], which may occur due to the autoregressive setup.
While the classification performance is clearly reported, the evaluation lacks standard ordinal regression metrics such as mean absolute error (MAE), which would better reflect the model’s ability to learn ordinal structure.
The method, while well written, introduces substantial complexity, and some components might benefit from a clearer intuitive explanation.
- 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.
- 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/2025/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 — could be 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 presents a creative and novel approach to DR grading by rethinking the task as an autoregressive ordinal regression problem, which is an original contribution to the field. The idea is well motivated, and the authors provide implementation details along with code, enhancing the impact and reproducibility of the work.
However, the model lacks a constraint mechanism to avoid generating invalid ordinal sequences, and the absence of ordinal regression metrics such as MAE limits the interpretability of the results in the context of ordinal prediction. These are important considerations that could be addressed in the rebuttal or future work.
Despite these concerns, the originality and clarity of the contribution justify a weak accept, as the method could inspire further developments in ordinal learning for medical imaging tasks.
- 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 AOR-DR, a novel autoregressive ordinal regression method that leverages clinical ordinal knowledge in DR grading. It decomposes grading into sequential steps, conditioning each on previous predictions and global image features. A diffusion process is employed to model conditional probabilities, enabling effective use of continuous image features without relying on patch-level context relearning.
- 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.
Reviewer found the following points is interesting:
- formulating DR Grading as auto-regressive sequence decoder, which helps to deal with the long-tailed distribution in DR data.
- the diffusion model is well-designed to integrate previous prediction with current input images through affine and cross attention module.
- the method archives good improvement compared other state-of-the-arts across four datasets.
- the paper is well-written and be easy to follow.
- 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.
-
Though using diffusion model inside ASD model achieves good performance, the motivation for using this approach is not well-motivated or validated by other alternative solutions. For e.g., what happen if we just use a simple transformer (can be pre-trained in advance) taking known predictions (y_{1:j-1}) global image feature and try to predict directly y_{j}? This should be investigated in the ablation study because diffusion model is one of key contributions of paper.
-
It’s unclear in Figure 3 why over Steps, the red regions (No-DR) is growing. Can you author explain this observation?
-
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My mainly concern is about the motivation of using diffusion rather than other options. Though, Reviewer still believes this paper is a solid contribution.
- 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 AOR-DR, a novel autoregressive ordinal regression framework for diabetic retinopathy (DR) grading. The method addresses two key challenges in DR classification: (1) long-tailed data distribution and (2) ambiguity in class boundaries. The authors decompose the ordinal regression task into sequential binary classification steps, where each prediction is conditioned on previous outputs and global image features. By integrating diffusion-based optimization, the model directly leverages continuous global features from a pre-trained ViT backbone without tokenization, enabling robust conditional probability modeling. Extensive experiments on four datasets demonstrate state-of-the-art performance, outperforming six recent ordinal regression 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.
- Autoregressive Label Sequence for Long-Tail Mitigation The autoregressive framework decomposes the ordinal regression task into sequential binary steps (e.g., predicting [1,1,1,0] for “Severe DR”), inherently addressing long-tailed data distribution. Tail classes (e.g., “Proliferative DR”) require more inference steps, forcing the model to focus on underrepresented categories during training. This design is validated by improved accuracy on tail classes.
- Diffusion-Based Robustness Enhancement The diffusion process models conditional probabilities in a continuous feature space (CLS token), avoiding the loss of global context caused by patch tokenization. By iteratively denoising predictions, the method reduces sensitivity to ambiguous class boundaries, leading to more stable and reliable predictions.
- 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.
- Lack of Hallucination Analysis The paper does not discuss potential hallucination risks in diffusion-based classification, such as generating invalid label sequences (e.g., [1,0,1,1] for an invalid ordinal transition). A robustness analysis against such failures or adversarial examples is missing.
- Slow DDPM Sampling Limits Clinical Applicability The inference relies on the original DDPM sampling (Eq. 6), which requires multiple iterative steps (e.g., 1000 steps). This introduces latency incompatible with real-time clinical deployment. Accelerated sampling methods (e.g., DDIM) are not explored, and no inference speed analysis is provided.
- Missing Ablation Against Non-Ordinal Baselines The experiments compare only with ordinal regression methods (e.g., CORAL, CORN). Non-ordinal approaches (e.g., standard classification with class-balancing losses) are excluded, making it unclear whether the gains stem from the autoregressive-diffusion architecture or ordinal modeling itself.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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/2025/en/REVIEWER-GUIDELINES.html
- Clarification on Affine Fusion and Cross-Attention Fusion The paper introduces two fusion strategies (Affine Fusion and Cross-Attention Fusion) but does not justify their selection or compare their advantages/disadvantages. For instance: Why are these two fusion mechanisms preferred over alternatives (e.g., concat)?A brief ablation study (e.g., accuracy/F1 differences between fusion strategies) would strengthen the methodology.
- Dataset Class Distribution Details The long-tailed distribution is a core challenge, but the paper lacks explicit statistics on the sample count per DR severity level (e.g., how many “Proliferative DR” cases exist in APTOS vs. DDR?). Including a table summarizing class distributions would help readers assess the severity of data imbalance and its impact on results.
- Incomplete Hyperparameter Specification Critical hyperparameters for the diffusion process，number of diffusion timesteps, training iterations) are not clearly stated. This limits reproducibility. Please provide these details in the “Implementation Details” section.
- Minor Grammatical Issues Example: “The model demonstrate superior performance…” → “The model demonstrates superior performance…” A thorough proofread is recommended to fix subject-verb agreement and tense inconsistencies.
- 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 — could be 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 introduces a novel and technically sound framework (AOR-DR) that effectively addresses long-tailed and ambiguous DR grading. The integration of diffusion models with autoregressive ordinal regression is innovative, and the extensive empirical validation across datasets demonstrates robustness. However, the study would benefit from: (1) additional ablation experiments to validate key design choices (particularly the diffusion model’s contribution versus simpler alternatives), and (2) a more thorough discussion of limitations, including potential hallucination risks in diffusion-based classification and practical constraints on clinical applicability (e.g., inference speed requirements for real-world deployment). While these aspects require further exploration, the methodological contribution remains significant and aligns well with MICCAI’s focus on advancing medical image analysis.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
Adding a discussion on prevention of non-acceptable sequence could improve the manuscript as suggested by two reviewers