List of Papers Browse by Subject Areas Author List
Abstract
Retinopathy of prematurity (ROP) staging via fundus imaging is critical for timely screening and treatment planning in premature infants. However, automated diagnosis is frequently impeded by low image quality and the high visual similarity between adjacent disease stages, both of which introduce significant diagnostic ambiguity. To address this, we propose ProtoEvi-ROP, an interpretable and uncertainty-aware framework designed to enhance the reliability of early-stage ROP staging.
Our approach synergizes a prototype-based reasoning mechanism for feature-level interpretability with evidential uncertainty modeling to explicitly quantify decision confidence. Additionally, we introduce a region-focused regularization strategy to stabilize prototype geometry, preventing feature drift.
Extensive experiments on two ROP fundus datasets demonstrate that ProtoEvi-ROP consistently outperforms state-of-the-art methods. Crucially, the estimated uncertainty serves as a robust reliability indicator, effectively filtering incorrect predictions. Selective prediction experiments reveal a clear accuracy-coverage trade-off, confirming that high uncertainty correlates with increased diagnostic risk, while decision space visualizations elucidate the synergy between interpretable prototype reasoning and uncertainty quantification.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2826_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{WuDon_Retinopathy_MICCAI2026,
author = { Wu, Donghan AND Shen, Wenyue AND Yuan, Lu AND Li, Heng AND Hao, Huaying AND Shi, Caiping AND Ye, Juan AND Zhao, Yitian},
title = { { Retinopathy of Prematurity Staging via Prototype-Evidential Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper addresses the diagnostic ambiguity in Retinopathy of Prematurity (ROP) staging caused by low image quality and high visual similarity between adjacent stages. It proposes a dual-stream framework named ProtoEvi-ROP. Its main contributions include: Integrating interpretable prototype learning with uncertainty-quantifying evidential learning. The prototype branch learns stage-specific morphological prototypes via distance metrics to handle “conflicting evidence” (features resembling multiple stages). The evidence branch models evidence strength via the Dirichlet distribution to handle “insufficient evidence” (ambiguous features). A “Region-Focusing Regularization” strategy is proposed, applying a central region mask to the prototype branch’s input to force the model to focus on peripheral vascular-avascular junction pathologies, stabilizing prototype learning. An improved Dempster-Shafer evidence theory rule is employed to fuse the belief masses from the two branches, better handling conflicts between them. Experiments on two ROP datasets show the method outperforms existing methods in metrics like accuracy. Furthermore, its estimated uncertainty effectively distinguishes correct from incorrect predictions, enabling risk-aware decision-making.
- 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.
Accurately identifies the core challenge in automated ROP staging—visual ambiguity and diagnostic uncertainty—and explicitly categorizes uncertainty into “conflicting evidence” and “insufficient evidence,” providing a clear motivation. The proposed ProtoEvi-ROP framework is well-designed, innovatively and systematically integrating prototype learning (providing geometric interpretability) and evidential learning (providing uncertainty quantification) within a single model. Conducts extensive experiments on two datasets, including comparisons with various SOTA methods, thorough ablation studies (validating the dual-branch and masking strategies), and calibration analysis of uncertainty (accuracy-coverage curve). Results show consistent performance gains. Goes beyond reporting performance numbers by providing in-depth analysis through feature space visualizations and uncertainty distribution statistics, explaining whythe model works and howits uncertainty estimates correlate with prediction errors, enhancing the work’s credibility.
- 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 core idea of combining “prototype learning” and “evidential learning” is not novel in medical image analysis or general machine learning. Prior works have explored integrating evidential uncertainty into prototype networks for open-set recognition or handling out-of-distribution samples. The paper fails to sufficiently argue the fundamental distinction and significant advantage of its “prototype-evidence” dual-stream design compared to existing hybrid approaches. The “novelty” lies more in the application task (ROP staging) and specific component design (e. g. , regional masking), not in a core paradigm shift. （2）The forced masking of the central 1/3 region as a “hard inductive bias” raises concerns about its generalizability. While ROP pathologies often occur in the periphery, the location and size of central structures like the optic disc are not absolutely constant across different devices or imaging conditions. This hard masking may discard potentially informative central features (e. g. , optic disc morphology, vascular arcade) and could harm model robustness with changing data distributions. The authors do not discuss the limitations of this design.
- 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
（1） In the introduction or related work section, please provide a more in-depth discussion differentiating your work from existing “prototype + uncertainty” hybrid approaches, explicitly stating the unique contributions of your work in terms of problem-driven (addressing specific types of diagnostic ambiguity) and technical implementation (e. g. , asymmetric masking, improved DS fusion rule). （2） Please supplement the discussion with the potential limitations and applicability boundaries of the regional masking strategy. Also, consider adding an analysis of computational complexity and inference time to assess the method’s practicality. （3） It is strongly advised to add comparative experiments with specialized uncertainty estimation and interpretability methods, particularly using uncertainty calibration metrics and prototype quality/interpretability for quantitative comparison. This would more powerfully demonstrate the dual advantages of the proposed framework.
- 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?
（1）Insufficient Novelty: The core of this paper is combining prototype learning and evidential learning. This idea itself is not novel, and the paper fails to sufficiently articulate its fundamental innovation compared to existing hybrid approaches. Its contribution lies more in the engineering implementation for the specific task of ROP (e.g., regional masking) rather than proposing a new methodology with broad impact. （2）The reliance on “Region-Focusing Regularization” as a strong prior lacks thorough validation and discussion regarding its generalizability and robustness. In real-world clinical settings with significant data variance, this hard masking could lead to information loss and performance degradation.
- 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 authors have carefully addressed the reviewers’ concerns and provided detailed clarifications regarding the motivation, architectural design, uncertainty modeling, clinical relevance, computational complexity, and dataset limitations of the proposed method. In particular, the revised manuscript now better explains the rationale behind the asymmetric uncertainty decoupling framework, the clinically motivated hard masking strategy, and the practical role of uncertainty-aware screening in reducing diagnostic risk.
The additional analyses on uncertainty mechanisms, computational efficiency, and ablation studies further strengthen the technical validity of the work. The authors also clarified the dataset construction process and committed to publicly releasing the dataset, which improves the reproducibility and potential impact of the study.
Overall, the manuscript presents a meaningful and technically sound contribution to uncertainty-aware ROP screening, with clear clinical significance and strong experimental support. Therefore, I recommend acceptance of the paper.
Review #2
- Please describe the contribution of the paper
This paper introduces ProtoEvi-ROP, an interpretable and uncertainty-aware framework tailored for the staging of Retinopathy of Prematurity (ROP). The core contribution lies in the integration of a prototype-based branch for feature-level interpretability and an evidential learning branch to quantify uncertainty. A coordination consistency loss is further employed to align the dual branches. Experimental results on two ROP datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches. The ablation studies and uncertainty analyses suggest that the framework provides a reliable mechanism for risk-aware clinical predictions.
- 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.
Methodological Novelty: The synergy between prototype-based learning and evidential deep learning is well-motivated, offering a dual perspective on both “why a decision was made” and “how confident the model is.” Clarity of Presentation: The methodology is logically structured and the technical components are well-articulated. Insightful Evaluation: The authors provide comprehensive analyses (e.g., Fig. 4) to substantiate the effectiveness of the uncertainty estimation and its alignment with interpretability, which is crucial for medical applications.
- 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.
Redundancy in Presentation: In Section 2.1 Evidential Branch part, the opening two sentences description appear redundant, as they essentially convey the same concept in slightly different phrasing. Parameter Sensitivity: The paper lacks a sensitivity analysis for several key hyperparameters, specifically the temperature \tau in Eq. (1) and the balancing coefficients \lambda_{evi} and \lambda_{cc} in Eq. (7). Exploring how these parameters influence model performance and uncertainty calibration would strengthen the technical depth. Clinical Impact of Uncertainty Filtering: While the authors show that filtering out high-uncertainty samples improves accuracy in Fig.4(d), this “rejection” mechanism may lead to a decrease in Recall, which is a critical metric in ROP screening to avoid missed diagnoses.The authors should discuss the clinical implications of these “rejected” cases.It would be highly beneficial to analyze the class distribution of these high-uncertainty samples. Are they predominantly from minority classes or borderline stages? Is there a correlation between model uncertainty and the difficulty levels perceived by human experts? An in-depth analysis of these limitation cases is needed to ensure the model’s reliability in a real-world clinical workflow.
- 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.
(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?
I recommend an Accept for this manuscript as it presents a well-engineered and conceptually sound framework, ProtoEvi-ROP, which effectively addresses the critical need for interpretability and reliability in medical AI. The integration of prototype-based feature learning with evidential uncertainty estimation is both technically innovative and clinically relevant. The authors have demonstrated superior performance across multiple ROP datasets, and the ablation studies provide strong evidence for the necessity of each component. While the sensitivity of certain hyperparameters and the potential impact of uncertainty-based filtering on Recall are areas for future discussion, the current work establishes a robust baseline for risk-aware clinical staging. The ability to quantify decision confidence significantly enhances the potential for human-AI collaboration in real-world screening workflows. Overall, this paper offers a complete and insightful contribution that is likely to be of high interest to the medical imaging community.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes ProtoEvi-ROP, a dual-stream framework for ROP staging that integrates:
Prototype-based learning for interpretable geometric reasoning Evidential deep learning for uncertainty estimation
The key novelty lies in:
A joint prototype evidential architecture to handle ambiguity in early-stage ROP A reduced Dempster Shafer fusion strategy to combine beliefs A region-focused masking strategy to stabilize prototype learning
The method aims to provide both accurate and trustworthy predictions for clinically challenging early-stage ROP classification.
- 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.
Well-motivated problem: Early-stage ROP ambiguity is clearly explained and clinically relevant. Combination of interpretability + uncertainty is meaningful and important in medical imaging. The dual-branch design (prototype + evidential) is conceptually sound and well justified. The use of Dempster–Shafer theory for fusion is interesting and adds theoretical grounding. Consistent improvements over SOTA across two datasets (Table 1 shows clear gains in accuracy, F1, AUC). Includes uncertainty analysis + selective prediction, which is valuable for clinical deployment. Ablation studies validate contributions of each component. The method provides interpretability via prototype space visualization, which is a strong practical advantage.
- 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 novelty at component level: The method mainly combines existing ideas: prototype learning evidential deep learning Dempster–Shafer fusion The novelty is mostly in integration rather than fundamentally new methodology.
2.Lack of comparison with recent uncertainty-aware medical models: The paper does not sufficiently compare with recent works on: Bayesian DL MC Dropout Deep ensembles which are standard baselines for uncertainty.
3.Dataset limitation: One dataset is private (PEFI-ROP), limiting reproducibility Total dataset size is relatively small (≈700–800 images each)
4.Generalization not validated: No cross-dataset evaluation No external validation
5.Prototype masking design is heuristic: The “central 1/3 masking” lacks strong justification or sensitivity analysis
6.Computational complexity not discussed: Dual-branch architecture likely increases cost but no analysis is provided
- 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
This is a well-structured and solid paper addressing an important clinical problem. The integration of interpretability and uncertainty is valuable, particularly for safety-critical applications like ROP screening.
However, the paper would benefit from:
stronger novelty justification beyond combining existing methods, broader comparisons with uncertainty modeling approaches, and more rigorous validation on external datasets.
Clarifying these aspects would significantly strengthen the 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.
(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 addresses an important problem and proposes a well-designed framework that combines interpretability and uncertainty estimation in a meaningful way. The experimental results are strong and consistent across datasets, and the analysis of uncertainty is particularly valuable for clinical applications.
However, the methodological novelty is somewhat limited, as the approach mainly integrates existing techniques rather than introducing fundamentally new components. Additionally, the evaluation lacks comparisons with standard uncertainty baselines and does not include external validation, which is important for medical imaging applications.
Overall, improvements in novelty justification and evaluation would strengthen it further.
- 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 authors addressed most of my major concerns in the rebuttal. In particular, they provided a clearer justification for the architectural novelty by explaining the asymmetric decomposition between prototype-based geometric ambiguity modeling and evidential reliability estimation. The clarification regarding the clinical motivation of the masking strategy and the uncertainty-driven human-in-the-loop workflow was also helpful.
The additional discussion on computational complexity and inference speed further strengthens the practical relevance of the method for real-world screening settings.
While some limitations still remain, particularly the lack of comparisons with standard uncertainty baselines such as MC Dropout and Deep Ensembles, as well as the absence of external validation, I believe the overall contribution is sufficiently strong for acceptance. The paper presents a well-structured and experimentally comprehensive framework with meaningful clinical motivation, strong quantitative results, uncertainty analysis, interpretability studies, and ablation experiments.
Overall the rebuttal adequately clarified the main concerns that motivated my initial weak reject score.
Author Feedback
1.Novelty（R1 & 3） Our framework is not a mere concatenation of existing tools, but a problem-driven architectural paradigm tailored to early-stage ROP. Standard prototypes overconfidently force geometric fits on weak evidence (Fig.3, Ablation II), while standard EDL lacks interpretability (Fig.3, Ablation I). Our fundamental innovation asymmetrically decouples uncertainty types. Mirroring clinical reasoning, our asymmetric architecture explicitly decouples functions into two parallel pathways: one pathway a morphological expert isolating conflicting evidence, and a global reliability gatekeeper isolating insufficient evidence. Rather than artificially averaging divergent predictions, our integration strategy explicitly preserves this inter-stream conflict. This transforms uncertainty into an actionable indicator of diagnostic risk, demonstrating a cohesive, structural innovation that resolves geometric confusion and evidence insufficiency far beyond a simple plug-and-play hybrid model
2.1/3 central hard mask （R1 & 3） The 1/3 spatial mask is not an arbitrary heuristic, but a principled anatomical prior based on ICROP-3 guidelines [3]. Clinically, ROP staging is strictly defined by morphological changes at the peripheral vascular-avascular junction. The central region (optic disc and macula) is highly salient but diagnostically irrelevant. Without the mask, prototype networks suffer from severe shortcut learning on these central features. Masking explicitly forces the prototype branch to align with clinical diagnostic criteria. This asymmetric design is empirically validated (Fig. 3, Ablation II), improving accuracy by 1.2% (SZ-ROP) and 1.3% (PEFI-ROP) over the unmasked baseline.
3.Computational Complexity (R1 & 3) ProtoEvi-ROP estimates uncertainty with a single deterministic forward pass, avoiding the N-fold inference overhead of MC Dropout and deep ensembles. Although slightly more costly than a single-stream baseline, it delivers improved performance. The final model (47.02M parameters, 8.26 GFLOPs) achieves 7.88 ms/image and 126.94 FPS on an RTX 5090 GPU. This favorable trade-off between performance, reliability, and efficiency makes the system ideal for real-time screening applications.
4.Uncertainty baselines (R1 & 3) We agree that MC Dropout and Deep Ensembles are standard uncertainty baselines. We have evaluated our internal uncertainty mechanisms in the current version of the manuscript (Fig. 3, Ablation I; Fig. 4c,d). We systematically compared standard softmax, prototype entropy, Dirichlet evidence, and our Reduced DS fusion, which consistently achieved the highest Accuracy, F1-score, and AUC. We will also make a comprehensive quantitative comparison against the suggested baselines in the camera-ready supplementary material.
5.Clinical Impact of Uncertainty Filtering & presentation (R2) We agree that maximizing Recall is critical in ROP screening. However, our uncertainty filtering does not discard samples, it serves as a Human-in-the-Loop triage mechanism. Forcing a deterministic prediction on highly ambiguous images risks dangerous false negatives. By flagging high-uncertainty cases for senior expert review, the system actively protects diagnostic safety. We will emphasize this workflow in the revision. Additionally, we thank the reviewer for identifying the textual redundancy in Section 2.1, which we will streamline for better flow.
6.Dataset limitation (R3) Acquiring high-quality fundus images from infants is notoriously difficult due to poor pupil dilation and eye movement. We respectfully clarify that the used datasets are substantial given the extreme challenges of neonatal ophthalmic imaging. Our study utilizes SZ-ROP (756 images), the largest publicly accessible ROP staging dataset, while our PEFI-ROP dataset (858 images) exceeds this benchmark. To resolve the concern about reproducibility, we formally commit to publicly releasing our dataset upon acceptance.
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 reviewers raise concerns that the methodological novelty is mainly integrative, combining prototype learning, evidential learning, and Dempster–Shafer fusion without sufficiently distinguishing the proposed design from prior uncertainty-aware prototype-based methods.
The authors should address the robustness and justification of key design choices, especially the central-region masking strategy, hyperparameter sensitivity, uncertainty filtering mechanism, and potential impact on recall and minority/borderline ROP stages.
The rebuttal should also clarify generalizability and reproducibility concerns by discussing dataset limitations, external or cross-dataset validation, comparisons with standard uncertainty baselines such as MC Dropout, Bayesian methods and deep ensembles, and computational complexity.
- 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 paper is accepted. The rebuttal addressed most major concerns, especially the motivation of the asymmetric uncertainty design, masking strategy, computational cost, and dataset limitations. Although the novelty is somewhat integration-oriented and external validation remains limited, the method is clinically meaningful, experimentally solid, and provides useful uncertainty-aware support for ROP screening.
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.
Overall, a well motivated paper, and the authors’ rebuttal addressed most reviewers’ concerns.
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 rebuttal substantially strengthened the case for acceptance. In particular, the authors clarified the motivation for the asymmetric dual-branch design, justified the central masking strategy using the clinical staging criteria, discussed computational efficiency, and explained the intended human-in-the-loop role of uncertainty filtering. They also committed to releasing the private dataset, which improves reproducibility and impact.