List of Papers Browse by Subject Areas Author List
Abstract
Diagnosing multiple diseases from retinal images is critical in real-world clinical applications, where patients often present with co-occurring conditions. However, existing approaches typically suffer from co-occurrence-induced lesion-disease mismatching: when diseases frequently co-occur, the model may spuriously associate lesions of one disease with another. Since all target diseases are present, this mismatch escapes supervision, leading to incorrect spatial reasoning and poor generalization when diseases appear independently during inference.
To address this challenge, we propose MD-PRSA, a novel framework for Multi-Disease diagnosis in retinal images via Patch-Level Reasoning and Selective Aggregation. MD-PRSA employs a patch-based transformer that computes disease-specific predictions for each patch using learned queries, and then aggregates these predictions through a multiple instance learning (MIL) mechanism. This selective aggregation enforces a form of spatial attribution constraint, ensuring that only patches with strong disease-specific evidence contribute to the final decision. This design naturally suppresses mismatched patches caused by co-occurring conditions and provides more robust and interpretable diagnoses under weak supervision.
We validate our method on two retinal image datasets with distinct imaging modalities: LID-FFA (fluorescein angiography) and ODIR-5K (color fundus photography). MD-PRSA achieves state-of-the-art results, improving mAUC by 0.84% on LID-FFA and 1.01% on ODIR-5K. Code will be made publicly available upon acceptance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0417_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
N/A
BibTex
@InProceedings{XieJia_MultiDisease_MICCAI2026,
author = { Xie, Jianyang AND Zhao, Yitian AND Meng, Yanda AND Zhang, Hao AND Alam, Uazman AND Zheng, Yalin},
title = { { Multi-Disease Diagnosis in Retinal Images via Patch-Level Reasoning and Selective Aggregation } },
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 proposes MD-PRSA, a framework for multi-disease retinal image diagnosis that aims to mitigate co-occurrence-induced lesion-disease mismatching. The method combines patch-level disease reasoning with selective Top-K aggregation under a MIL-style formulation. In this case, this selective aggregation enforces a form of spatial attribution constraint, ensuring that only patches with strong disease-specific evidence contribute to the final decision, which naturally suppresses mismatched patches caused by co-occurring conditions and provides more robust and interpretable diagnoses under weak supervision. Experiments on LID-FFA and ODIR-5K show consistent improvements over several baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(1) The paper addresses an important and practically relevant problem. In real clinical settings, multiple retinal diseases often co-occur, and the discussion of lesion–disease mismatching is well motivated. (2) The proposed method is conceptually clear and easy to follow. The combination of disease-specific patch-level reasoning with selective aggregation is intuitive and well aligned with the problem setting. (3) The experimental results are generally solid. The proposed method shows consistent improvements over strong baselines across two datasets with different imaging modalities. In particular, the conditional evaluation in Section 3.4 is well designed, as it directly supports the central claim of reducing spurious associations under disease co-occurrence.
- 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 appears to be somewhat sensitive to the choice of Top-K, and the selection strategy could be further clarified and better justified. (2) While the evidence for improved interpretability and robustness to lesion–disease mismatching is encouraging, it remains somewhat preliminary. Providing quantitative evidence that the selected patches correspond to clinically meaningful disease regions would make this claim more convincing.
- 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?
Important clinical values, clear method description and sufficient experiments.
- 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.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes MD-PRSA, a multi-disease retinal diagnosis framework that generates patch-level disease-specific predictions via transformer decoder with learned disease queries, then aggregates them through top-K MIL pooling to suppress co-occurrence-induced lesion-disease mismatching.
- 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.Well-motivated problem 2.Sound experimental design 3.Consistent improvements across two modalities 4.Interpretability via patch-level attribution
- 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 methodological novelty. The pipeline (CNN patch encoder → transformer decoder with disease-specific queries → patch-level prediction → top-K MIL aggregation) assembles well-established components. Patch-level prediction with MIL aggregation is a mature paradigm in computational pathology (e.g., CLAM [Lu et al., 2021], DSMIL [Li et al., CVPR 2021], TransMIL [Shao et al., NeurIPS 2021]). The disease-specific query mechanism closely resembles Q2L [14], which the authors acknowledge; the primary difference is replacing global pooling with top-K selection. This substitution, while reasonable, constitutes incremental modification rather than a conceptual advance.
2.Marginal quantitative improvements without statistical validation. The reported mAUC gains are +0.84% on LID-FFA and +1.01% on ODIR-5K. Improvements of this magnitude can fall within random variance across training runs. No confidence intervals, standard deviations, or statistical significance tests are reported. Without such validation, it is unclear whether the improvements are reliable or merely stochastic.
3.Only two datasets are used, both from the retinal imaging domain. No experiments on other multi-label medical imaging benchmarks (e.g., chest X-ray with CheXpert/MIMIC-CXR, or dermatology) are provided to support the claimed generalizability. Additionally, all experiments use a single backbone (DenseNet121), which is not a commonly-evaluated backbone. It would be useful to clarify the selection reason, such as following the previous works or benchmarks. Testing on diverse architectures (e.g., ResNet, ViT) would strengthen the generalization claim.
4.The “first unified framework” claim is not well-supported. The paper states “To our knowledge, MD-PRSA is the first unified framework that combines disease localization with disease-specific feature modeling for multi-disease retinal diagnosis.” However, prior works such as CABNet [7] (cited by the authors) and attention-guided methods already combine spatial localization with disease-specific classification. This claim needs more careful qualification or stronger evidence of distinction.
5.Insufficient analysis of the top-K hyperparameter. The top-K selection is central to the MIL aggregation, yet K=10 for LID-FFA and K=50 for ODIR-5K are used with limited justification. The ablation (Table 2) shows sensitivity to K, but does not explain why optimal K differs by 5× across datasets or how to select K for a new dataset without exhaustive search. A principled or adaptive selection mechanism would strengthen the approach.
Others: The introduction contains a typo (“In summury”)
- 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?
See weakness.
- 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.
I appreciate the authors’ response which have addressed all of my concerns.
Review #3
- Please describe the contribution of the paper
The authors propose MD-PRSA, a novel framework designed to tackle the specific challenge of co-occurrence-induced lesion-disease mismatching in multi-disease retinal diagnosis. The primary idea is the integration of a Transformer-based Patch-Level Reasoning module with a Multiple Instance Learning (MIL)-based Selective Aggregation mechanism. This design enforces a spatial attribution constraint, forcing the model to identify a minimal subset of causally relevant patches for diagnosis rather than relying on spurious correlations caused by frequently co-occurring diseases. The method demonstrates state-of-the-art performance on two distinct retinal datasets (LID-FFA and ODIR-5K).
- 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 identifies and explicitly defines a critical pathology in multi-disease diagnosis: “co-occurrence-induced lesion-disease mismatching.” 2.The method achieves State-of-the-Art (SOTA) results on two challenging datasets with different modalities.
- 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 comparison methods are mostly common methods in multi-label domains, lacking specialized methods for the field of retinal diseases. 2.The performance of the model is highly dependent on the choice of K (the number of top patches selected). While the authors discuss this in the ablation study, the optimal K differs significantly between datasets (K=10 vs K=50). 3.For high-resolution retinal images, the Transformer decoder with cross-attention could lead to significant computational overhead compared to simpler CNN baselines. Its computational complexity shoud be discussed.
- 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.
(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?
Please reffer to the major weaknesses.
- 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 thank all reviewers for their constructive feedback and positive recognition of the clinical importance of co-occurrence-induced lesion–disease mismatching, the interpretability of the proposed framework, and the overall experimental design. We address the main concerns below. 1.Clarification of the novelty (R2) The novelty of this work lies not in introducing entirely new architectural components, but in formulating and addressing a previously under-discussed problem in multi-disease retinal diagnosis: co-occurrence-induced lesion–disease mismatching, directly evaluated through the conditional evaluation protocol in Sec.3.4 and Fig.3.Existing MIL-based methods (e.g., CLAM, DSMIL, TransMIL) are primarily designed for binary or multi-class bag classification, where identifying discriminative instances is sufficient for bag-level prediction. In contrast, multi-label retinal diagnosis involves multiple co-occurring diseases under only image-level supervision, leaving disease-specific lesion attribution fundamentally under-constrained. Consequently, models may achieve correct predictions while relying on lesions from other co-occurring diseases. While Q2L extends transformer-based learning to multi-label classification, it does not explicitly address lesion–disease attribution ambiguity under disease co-occurrence. Our framework instead combines disease-specific patch reasoning with selective aggregation to encourage predictions to rely on disease-specific lesion evidence. 2.Performance gains and robustness (R2) We acknowledge the reviewer’s concern regarding statistical variability across training runs. Nevertheless, the observed improvements show a consistent pattern across retinal imaging modalities. Importantly, the improvements are not limited to aggregate performance. As shown in Sec.3.4 and Fig.3, MD-PRSA consistently improves conditional AUC across co-occurring disease pairs, suggesting reduced reliance on spurious co-occurrence correlations and improved disease-specific reasoning. 3.Top-K selection (R1, R2, R3) The optimal K varies with disease morphology and imaging modality. As discussed in Sec.3.5, ODIR-5K contains diseases with distributed retinal manifestations, whereas LID-FFA mainly involves localized vascular abnormalities, naturally favoring different aggregation scales. This demonstrates that selective aggregation can adapt to disease spatial characteristics. This consideration is also discussed in the manuscript. 4.“first unified framework” claim (R2) We agree that the phrase “first unified framework” may be interpreted too broadly. We intended to highlight a key methodological distinction: prior methods, such as CABNet and other attention-guided approaches, primarily compute attention at the feature level, whereas our framework performs explicit disease-specific patch selection at the logit level through selective aggregation. We will revise the wording to avoid overstatement. 5.Dataset diversity and backbone selection (R2, R3) Our work specifically focuses on multi-disease retinal diagnosis, as reflected in the paper title. DenseNet121 was adopted to ensure fair comparison across all baselines and is widely used in retinal image analysis due to its parameter efficiency and stable optimization behavior. While broader evaluation across additional domains and architectures may further strengthen generalizability, the current experiments consistently support our central contribution in mitigating lesion–disease mismatching across diverse retinal imaging settings. (R3) The comparisons were also not limited to generic multi-label methods. Retinal-specific approaches such as Q2L Causal and MCAR were included under the same experimental setting. 6.Interpretability and computational cost (R1, R3) Fig.4 shows that the selected patches consistently correspond to clinically relevant lesion regions. In addition, the decoder operates on patch tokens, making the approach computationally practical.
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 appreciate the paper’s clinical motivation in addressing co-occurrence-induced lesion-disease mismatching and acknowledge the solid performance improvements. However, the authors must comprehensively address all the critical concerns raised by reviewers during the rebuttal.
- 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 authors have addressed most of the concerns raised by reviewers. I recommend Accept.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper addresses lesion–disease mismatching under multi-disease co-occurrence in retinal image classification with a conceptually clear framework combining disease-specific patch-level reasoning and selective top-K aggregation. It demonstrates consistent improvements over strong baselines on two datasets, and presents adequate analysis and discussion points.
R2’s initial concern regarding limited methodological novelty was satisfactorily addressed in the rebuttal, prompting a recommendation change from weak reject to accept. R3’s concerns regarding retinal-specific baselines, sensitivity to K, and computational overhead are, in my assessment, also adequately addressed, though R3 did not submit a final recommendation. I recommend acceptance.
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 paper addresses a clinically relevant and underexplored problem in multi-disease retinal diagnosis: lesion–disease mismatching under disease co-occurrence. This problem formulation is meaningful, and the proposed patch-level reasoning with selective aggregation is well aligned with it. The method is clearly presented, shows consistent improvements on two retinal datasets with different modalities, and the conditional evaluation directly supports the paper’s main claim.
The rebuttal addressed the main concerns satisfactorily. In particular, it clarified that the contribution is primarily problem-driven rather than based on entirely new architectural blocks, and it better positioned the method relative to prior MIL and multi-label approaches. The explanation of the Top-K behavior across datasets is reasonable, and the authors also agreed to moderate the potentially overstated “first unified framework” claim.
While some concerns remain about limited architectural novelty, sensitivity to Top-K, and the absence of statistical validation across runs, these are not strong enough to outweigh the paper’s strengths.