List of Papers Browse by Subject Areas Author List
Abstract
Age-related Macular Degeneration exhibits heterogeneous lesion morphologies and spatial distributions across global fundus and local macula region. Previous studies compress disparate lesion features into a single scale representation, leading to representational bottleneck which dilutes peripheral features and allows extensive lesions to dominate subtle ones. To address this challenge, we propose a Collaborative Multi-scale Representation (CoMRep) learning. Dual-branch feature extractor produces multiscale features for global fundus and local macula. The subsequent Global-local Cross-scale Attention projects each feature into multiple latent representations, yielding diverse perspectives of lesion features with cross-scale attention. Collaborative Glocal Feature Aggregation utilizes perspective experts to refine diverse glocal features and intergrate for final prediction. Experiments on the ADAM dataset demonstrates that the proposed CoMRep exceeds 0.73% AUC, 5.10% accuracy, and 3.19% recall compared to the previous state-of-the-art method with the same experimental setting. Qualitative evaluation further indicates that CoMReP captures heterogeneous AMD lesions with multiple glocal feature mixture. The code is available at https://github.com/jny0812/CoMRep.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4365_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/jny0812/CoMRep
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JanNay_Collaborative_MICCAI2026,
author = { Jang, Nayeon AND Ko, Seonghyeon AND Bum, Junghyun AND Le, Duc-Tai AND Son, Chang-Hwan AND Choo, Hyunseung},
title = { { Collaborative Multiscale Representation Learning for AMD Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces CoMRep, a collaborative multi-scale representation learning framework designed for the robust detection of Age-related Macular Degeneration (AMD). The core innovation lies in its explicit handling of lesion heterogeneity by decomposing fundus images into global and local (macular) streams, thereby overcoming the “representational bottleneck” typical of single-scale approaches. Key Contributions Pioneering Multi-scale Formulation: To the best of the authors’ knowledge, this is the first study to formalize AMD detection as a multi-scale learning problem. By jointly modeling the global fundus context and fine-grained macular details, the framework prevents dominant lesions from overshadowing subtle pathological features. Global-Local Cross-scale Attention (GCA): The authors propose a novel GCA mechanism that projects features into multiple latent subspaces. This allows for bidirectional interaction between global and local scales, generating diverse “glocal” perspectives that capture complex lesion patterns. Collaborative Glocal Feature Aggregation (CGFA): The architecture incorporates a Mixture-of-Experts (MoE) based aggregation module. This module utilizes a learned Top-k router to selectively activate “perspective experts,” specializing the model’s ability to represent heterogeneous lesion characteristics efficiently. Validated Empirical Excellence: CoMRep demonstrates significant performance gains on the ADAM dataset, outperforming previous state-of-the-art methods by 5.10% in accuracy and 3.19% in recall. Notably, it achieves these results without the need for additional training data or high-cost, dense pixel-level annotations.
- 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.
Clinically grounded problem formulation. It effectively reframes AMD severity grading as a multi-scale learning challenge, moving beyond prior approaches that treat heterogeneous lesions with a single-scale representation. This shift acknowledges the clinical reality that AMD manifestations vary widely in size, location, and morphology—ranging from macula-centered to peripheral patterns. Logically coherent architecture. The proposed framework is thoughtfully structured as an integrated pipeline rather than an ad hoc assembly of modules: A dual-branch backbone extracts both global fundus features and localized macular ROI features. A cross-scale attention mechanism (GCA) enables bidirectional interaction through multi-latent representations. A mixture-of-experts module (CGFA) selectively aggregates information based on lesion characteristics. Each component directly addresses a specific limitation identified in existing methods, resulting in a well-justified design. Innovative use of cross-scale attention. Instead of simple feature concatenation or averaging, the GCA module learns diverse latent “views” of lesion features through bidirectional attention between global and local branches. This approach is particularly suited to medical images where multiple coexisting pathologies require nuanced representation. Effective modeling of heterogeneity via MoE. The CGFA module employs Top-k expert routing to dynamically adapt to different lesion patterns. This application of mixture-of-experts helps prevent representation collapse and enhances model flexibility—offering a novel and appropriate solution for handling AMD’s diverse phenotypic expressions. Demonstrated empirical gains under fair conditions. On the ADAM dataset, the method achieves consistent improvements—+5.10% in accuracy and +3.19% in recall—while maintaining competitive AUC, all without using additional training data. These results support the claim of data-efficient learning, a critical advantage in medical imaging contexts where labeled data is scarce. Rigorous validation of design choices. Comprehensive ablation studies isolate the contributions of GCA, CGFA, and the dual-branch structure, while expert routing analysis reveals interpretable specialization patterns. This evidence strengthens confidence that performance gains stem from intentional architectural decisions. Strong clinical alignment. From macular ROI extraction to the focus on recall (a key metric for screening), the methodology reflects domain knowledge and practical clinical needs, enhancing its potential for real-world deployment in AMD screening programs.
- 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.
Incremental Methodological Novelty: The framework effectively integrates several established concepts, such as dual-branch architectures, cross-attention, and Mixture-of-Experts (MoE) with Top-k routing, but lacks a fundamentally new theoretical contribution. While the integration is technically sound, the claim of being the “first” to formulate AMD detection as a multi-scale problem appears overstated, as global-local ROI approaches are well-documented in retinal image analysis. Concerns Regarding Generalization and Scale: The evaluation is confined to the ADAM dataset (1,200 images) from a single clinical source. The absence of external validation or cross-dataset testing makes it difficult to assess the model’s robustness against variations in imaging hardware, population demographics, and acquisition protocols. Marginal Improvements in Discriminative Power: While gains in accuracy and recall are noted, the improvement in AUC is relatively modest (~0.73%). Given the increased architectural complexity introduced by the MoE and dual-branch components, the trade-off between performance gains and computational overhead warrants a more critical justification. Dependency on Spatial Priors (Fovea Coordinates): The model’s reliance on pre-defined fovea coordinates for ROI extraction limits its “end-to-end” utility. The authors do not address the potential for error propagation if fovea localization fails, nor do they discuss the additional annotation burden this requirement imposes for real-world deployment. Insufficient Benchmarking Against Diverse Paradigms: The comparative analysis is somewhat narrow. The paper would benefit from benchmarking against recent fundus foundation models or segmentation-based pipelines that explicitly localizes lesions, rather than comparing primarily against standard CNN/Transformer backbones. Limited Interpretability and Clinical Depth: Despite the clinical motivation, the paper lacks a rigorous analysis of how the MoE routing correlates with specific lesion types (e. g. , drusen vs. atrophy). Furthermore, the study stops at binary detection without exploring clinical endpoints like disease staging or progression risk.
- 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?
1.Practical Clinical Impact (Strength) The paper addresses an important limitation in retinal image analysis, namely the representational bottleneck where global feature aggregation tends to suppress subtle, localized macular abnormalities. The proposed CoMRep framework effectively tackles this issue by explicitly modeling both global fundus context and local macular information. The reported improvements on the ADAM dataset—particularly the +5.10% accuracy and +3.19% recall—are meaningful. In a screening setting, even modest gains in recall can translate into a tangible reduction in missed AMD cases, highlighting the practical relevance of the approach. 2.Well-motivated and Cohesive Architecture (Strength) Although the individual components (dual-branch feature extraction, cross-attention, and mixture-of-experts) are not novel in isolation, their integration is thoughtfully designed and well-aligned with the problem. The Global-local Cross-scale Attention (GCA) module enables effective bidirectional interaction between global and local representations, which is particularly suitable for modeling heterogeneous lesion distributions. The use of a Top-k routing strategy in the aggregation stage is also a pragmatic design choice, balancing representational flexibility with computational efficiency—an important consideration for real-world deployment. 3.Limited Methodological Novelty (Weakness) The main limitation of the work lies in its incremental nature. The proposed framework primarily combines existing techniques rather than introducing fundamentally new modeling paradigms. Core components such as transformer-based cross-attention and mixture-of-experts are well-established, and their adaptation to this task, while effective, does not constitute a major conceptual advance. Furthermore, the claim of being the “first” to formulate AMD detection as a multi-scale problem appears somewhat overstated, as prior work has explored ROI-based or multi-branch strategies in fundus analysis. 4.Limited Validation and Generalization Analysis (Weakness) The experimental validation is restricted to a single dataset (ADAM), which raises concerns about generalizability. Without evaluation on external datasets from different institutions or imaging devices, it is difficult to assess robustness to domain shift. In addition, the reliance on fovea coordinates for macular ROI extraction introduces an implicit dependency on accurate spatial priors, which may not always be available or reliable in fully automated pipelines. The paper would benefit from further analysis of these aspects. Conclusion Overall, this is a solid and well-executed application-driven study. The method is carefully designed and demonstrates clear empirical improvements on a clinically relevant task. However, the contribution is primarily incremental, and the evaluation scope is somewhat limited. As such, the paper sits slightly above the acceptance threshold: it offers practical value and technical soundness, but lacks the level of novelty or comprehensive validation typically associated with stronger accept decisions.
- 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
Review #2
- Please describe the contribution of the paper
The paper proposes CoMRep, a novel framework designed to address lesion heterogeneity in Age-related Macular Degeneration (AMD). A key strength of the study is the formalization of a “Glocal” dual-branch strategy that integrates features from the global fundus and the local macula, effectively replicating the diagnostic approach of a retinal specialist.
The architecture is driven by the Global-local Cross-scale Attention (GCA), which facilitates bidirectional interactions between scales. This works in tandem with the Collaborative Glocal Feature Aggregation (CGFA) module to manage the structural diversity of AMD. By employing a Mixture of Experts rather than a conventional classifier, the CGFA allows the model to specialize its analysis based on specific lesion patterns.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clinically-Inspired Formulation: The study’s primary contribution is the treatment of AMD detection as a multi-scale problem. By linking global fundus anatomy with local macular details, the framework replicates the hierarchical diagnostic workflow used by clinical specialists.
2.Bidirectional Feature Synergy: The Global-local Cross-scale Attention (GCA) facilitates a sophisticated exchange of information. It allows the global pathway to pinpoint areas requiring high-resolution evidence while providing the local pathway with the necessary spatial context to interpret fine-grained features accurately.
3.Adaptive Feature Aggregation: The Collaborative Glocal Feature Aggregation (CGFA) provides a robust solution to lesion heterogeneity. Through the use of a Mixture of Experts (MoE) and a Top-k Router, the architecture selectively activates specialized “perspective experts,” ensuring the model is uniquely tuned to the diverse structural characteristics of AMD lesions.
- 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.Dependency on Manual Fovea Localization: A significant bottleneck in the proposed CoMRep framework is the requirement for annotated fovea coordinates to extract the macular RoI. This reliance on external spatial metadata restricts the model’s autonomy and limits its deployment in high-throughput screening environments where manual pre-processing is impractical. Integrating a fovea localization sub-network would be necessary to achieve a truly automated diagnostic tool.
2.Restricted Domain Generalizability: The evaluation of the model is currently confined to a single dataset, which raises concerns regarding its robustness against domain shift. Medical imaging often suffers from high variance across different hardware scanners and acquisition protocols. Without validation on multi-center or multi-vendor datasets, the generalizability of the “Glocal” features to unseen distributions remains unproven.
- 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
Yes, for equation number 10 you did not mention the value of λ in all 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The most compelling factor is the model’s mimicry of human diagnostic patterns. By formalizing a dual-branch strategy that processes both global fundus context and local macular details, the authors move away from “black-box” processing toward a biologically inspired framework. Furthermore, The technical contribution is robust, particularly through the introduction of the Global-local Cross-scale Attention (GCA) and Collaborative Glocal Feature Aggregation (CGFA).
- 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 presents Collaborative Multiscale Representation (CoMRep) learning for Age-related Macular Degeneration (AMD) detection. It utilizes a dual-branch feature extractor to process the global fundus image and a cropped local macular region independently. The extracted multiscale features are refined via a Global-local Cross-scale Attention (GCA) module and aggregated using Collaborative Glocal Feature Aggregation (CGFA), which leverages a Mixture-of-Experts routing strategy to dynamically accommodate heterogeneous lesion characteristics.
- 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 paper accurately identifies that AMD lesions are highly diverse in morphology and spatial distribution, often extending beyond the macula. The dual-branch approach directly and logically addresses the representational bottleneck inherent in single-scale compression methods.
- 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 method is exclusively evaluated on the ADAM dataset. Given the high complexity of the proposed architecture (GCA + MoE), it is difficult to assess its generalizability and robustness to other cohorts, varying disease prevalences, or different imaging devices without external validation. (2). The local feature extraction pathway relies on a provided fovea coordinate to crop the macular Region of Interest. While fovea localization is a complex and distinct task in its own right, and it is entirely reasonable that the authors focused primarily on the classification architecture, this dependency means the current pipeline is not fully end-to-end for raw clinical deployment. The paper would be significantly strengthened by a brief discussion on how this framework might be coupled with an automated fovea detector, or ideally, a small robustness analysis showing how sensitive the CoMRep model is to slight coordinate perturbations (simulating errors from an upstream automated localizer). (3). The comparative baselines in Table 1 largely consist of generic classification backbones (ResNet, SwinT, ViT) and standard ADAM Challenge entries. The paper lacks comparisons with more recent state-of-the-art multi-scale or dual-branch architectures specifically designed for retinal imaging. (4). The provided open-source code link is not accessible, and the manuscript contains multiple typographical and other errors.
- 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 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.
(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?
Multi-latent projection and MoE/Top-k expert routing are both fairly standard modeling patterns. While combining these modules for AMD detection is reasonable from an engineering perspective, the methodological novelty is limited. There are also multiple issues in the formulation on pages 4–5.First, the GCA section defines 2M latent representations, but the subsequent cross-attention “slots” are fixed to four. The t-SNE visualization and expert-routing heatmap presented in the paper only show that features are distributed across different subspaces, or that different tokens exhibit different expert-selection frequencies. However, this does not directly demonstrate that the model captures heterogeneous AMD lesions, nor does it show that the model truly focuses on clinically relevant lesion regions.
- 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
We sincerely thank the reviewers and area chair for positive responses and early acceptance. We will revise the camera-ready version following your comments.
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.
The paper presents a well-motivated approach and has received generally positive feedback from the reviewers. The proposed method is technically sound and shows promising performance, and I agree that it provides a meaningful contribution to the field.
That said, several important aspects could be further strengthened in the final version. The most notable concern is the limited evaluation setting. The experiments are conducted on a single dataset (ADAM), which makes it difficult to assess the robustness of the proposed “Glocal” features under domain shift. Given the variability in medical imaging across devices and institutions, it would be beneficial for the authors to better justify or discuss the expected generalizability of the method.
In addition, the method relies on fovea-based spatial priors for ROI extraction, which introduces a dependency on accurate anatomical localization. The robustness of the approach to potential errors in fovea localization, especially in fully automated or real-world settings, should be further discussed or analyzed.
Finally, the scope of baseline comparisons could be clarified. While the current comparisons against generic architectures and challenge entries are reasonable, the paper would be further strengthened by discussing its relationship to more recent task-specific designs (e.g., multi-scale or dual-branch retinal models), or by providing a rationale for the current choice of baselines.
Overall, despite the above limitations, the paper presents a solid contribution with practical relevance. I recommend acceptance, while encouraging the authors to address the above points to further improve the clarity and impact of the work.