Abstract

Large-scale pretrained foundation models have revolutionized general medical screening, but often falter on rare diseases because such conditions are underrepresented in real-world clinical datasets. While retrieval-augmented diagnosis attempts to mitigate this, conventional static methods frequently succumb to the hubness problem, retrieving visually similar but semantically incorrect common diseases. To address this, we propose Evo-RAD, a self-evolving agentic framework that transforms evidence acquisition into a dynamic decision-making task. We formulate retrieval as a Markov Decision Process (MDP) where a graph-based agent observes the reference set state and executes actions to purge discordant evidence (DELETE), acquire pathologically consistent samples (INSERT), or conclude the evolution (TERMINATE). Optimized via Group Relative Policy Optimization (GRPO) with a homogeneity-aware reward,the agent learns to maximize the diagnostic homogeneity of the support reference set. Experiments on retinal disease benchmarks show that Evo-RAD substantially improves rare-disease diagnosis, outperforming retinal foundation models by +21.04%, while also surpassing retrieval-based and parameter-efficient fine-tuning methods by +3.56%. Code is available at https://github.com/SDH-Lab/Evo-RAD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1081_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1081_supp.zip

Link to the Code Repository

https://github.com/SDH-Lab/Evo-RAD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{XiaWan_EvoRAD_MICCAI2026,
        author = { Xia, Wangding AND Du, Ye AND Lin, Jiashi AND Wang, Meng AND Shi, Danli AND Wang, Shujun},
        title = { { Evo-RAD: Navigating Rare Retinal Disease Diagnosis via Self-Evolving Agentic Retrieval } },
        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

    The paper proposes Evo-RAD, a self-evolving agentic retrieval-augmented diagnosis framework for rare disease diagnosis, reforming it as a dynamic decision-making task. It formulates retrieval as a Markov Decision Process where a graph-based policy agent iteratively refines the reference set through insertion, deletion, and termination actions. The agent is trained using Group Relative Policy Optimization with a homogeneity-aware reward to maximize diagnostic coherence, demonstrating improved performance over foundation models and parameter-efficient fine-tuning methods on curated datasets.

  • 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 formulation of retrieval as a sequential Markov Decision Process with discrete actions is a novel and interesting alternative to traditional static nearest-neighbor retrieval. The motivation behind this formulation is well justified by the so-called hubness problem in rare disease diagnosis.

    • The proposed homogeneity-aware reward is validated through the ablation studies, showing the individual contributions of the trajectory-level and step-level reward components.

    • The visualization demo provided in the supplementary material effectively illustrates the iterative reference set refinement process, offering intuitive clarification of the proposed framework.

  • 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 evaluation is restricted to custom splits derived from a single clinical dataset, raising questions about whether the dataset and findings are generalizable.

    • There lacks a practicality analysis regarding the computational cost and latency introduced by the iterative graph updates and foundation model dependency, particularly compared with static retrieval and foundation model baselines.

  • 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

    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?

    The proposed framework is thorough, and its motivation is well-justified within the medical domain context. While concerns regarding generalizability and computational cost remain, the publicly available repository relieves some of them.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 sufficiently addressed the main concerns regarding generalizability and efficiency of the proposed framework.



Review #2

  • Please describe the contribution of the paper

    This paper proposes Evo-RAD, an agentic retrieval framework for rare retinal disease diagnosis. Its main contribution is to move beyond static nearest-neighbor retrieval by modeling evidence selection as a sequential decision process, in which a graph-based policy iteratively removes inconsistent references, inserts more relevant cases, and stops when the retrieved evidence becomes sufficiently coherent. The paper further introduces a homogeneity-aware reward and GRPO-based training scheme, and demonstrates improved performance over foundation-model, static-retrieval, and PEFT 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.

    The novelty lies in reformulating retrieval-augmented diagnosis as a sequential evidence refinement process, rather than relying on static nearest-neighbor retrieval.

  • 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.

    I highly recommend the authors address a few points regarding the initial retrieval stage. Specifically, the paper is missing a natural intermediate baseline: an initial Top-K retrieval followed by a trainable re-ranking module, such as a RefineNet. If there happens to be at least one correct rare-disease sample within the initial K=8 candidates, could a well-trained re-ranking network identify it and yield a correct diagnosis? Without this baseline, it is difficult to determine whether the performance gain comes from the sophisticated multi-step agent architecture or merely from introducing a learnable scoring mechanism over the retrieved set. Additionally, the paper lacks direct quantitative data, such as Recall@K or Hits@K, for the initial reference set to empirically support the hubness problem claim. If the ground-truth rare disease samples are frequently already present in the initial Top-8, the mechanism of fetching new candidates from the external buffer might be considered over-engineered. Conversely, showing that the initial Recall@8 is critically low would serve as the ultimate justification for why the agent must pull in new evidence from the broader pool. These questions are not meant to dismiss the value of the agentic approach, which I find highly compelling, but rather to help the authors theoretically and empirically rule out simpler alternatives. I look forward to reading the authors’ response and seeing this excellent work become even stronger.

  • 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

    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?

    I give this paper a weak accept because it addresses an important and difficult problem, proposes a clearly novel formulation, and reports strong empirical results. In particular, the idea of evolving the retrieved evidence set through sequential DELETE/INSERT/TERMINATE decisions is interesting and appears to be a meaningful step beyond static retrieval. The paper is well written and the overall technical contribution is solid.

    My main reservation is that the necessity of the full agentic formulation is not yet completely established. The paper lacks comparison to a simpler reranking/refinement baseline on top of the same initial retrieval set, and it does not report initial retrieval quality metrics that would more directly justify the need for iterative evidence evolution. For this reason, I view the work as accept-worthy but not yet strong enough for a higher score.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    Authors have addressed my concern.



Review #3

  • Please describe the contribution of the paper

    The paper’s main contribution is to reformulate retrieval-augmented diagnosis for rare retinal diseases as a multi-step evidence refinement problem rather than a one-shot nearest-neighbor lookup. The key motivation is that conventional retrieval often returns visually similar but pathologically/semantically mismatched reference cases, especially under rare-disease class imbalance and hubness, which can bias diagnosis toward common diseases. To address this, the authors propose Evo-RAD, an agentic retrieval framework that models reference-set construction as a Markov Decision Process. Starting from an initial top-K retrieved set, a graph-based policy iteratively performs DELETE, INSERT, and TERMINATE actions to remove discordant references, add more diagnostically consistent cases, and stop when the support set becomes sufficiently coherent. The policy is trained with GRPO and a homogeneity-aware reward that encourages correct diagnosis, increased label purity, and stronger semantic consistency in the evolved evidence set. In this way, the paper’s core contribution is not just a new retrieval module, but a shift from static similarity-based retrieval to dynamic, decision-driven evidence search for rare disease diagnosis, which leads to improved performance on retinal benchmarks.

  • 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 a concrete failure mode of static retrieval for rare retinal disease diagnosis—hubness causes retrieval of visually similar but pathologically mismatched common-disease references—and builds the entire method around fixing that problem; (2) it introduces a specific and interesting formulation, casting retrieval refinement as an MDP with DELETE, INSERT, and TERMINATE actions instead of a one-shot top-K lookup; (3) the graph state is a meaningful design choice because it models both query relevance and agreement among retrieved references, rather than treating cases independently; (4) the homogeneity-aware reward is well aligned with the task, explicitly optimizing diagnosis accuracy, label purity improvement, and semantic density of the support set; and (5) the method shows sizable gains over strong retinal FM, static retrieval, and PEFT baselines, with ablations showing that the state design and reward components materially contribute to performance.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Major weaknesses: (1) the paper has limited clinical validation: evaluation is entirely retrospective on curated benchmarks from Retina Image Bank, with no external-hospital test set, reader study, workflow study, or prospective evidence of clinical feasibility; (2) the method’s final diagnosis is still a majority vote over retrieved labels, so the overall system remains heavily dependent on the quality and label correctness of the retrieval database, and the paper does not test robustness to noisy labels or open-world cases; (3) the training objective is quite strongly supervised, since trajectory reward and step rewards explicitly depend on the ground-truth label and whether inserted/deleted samples are correct, which may make the setting less realistic than the inference-time problem; (4) the method also depends on LLM-based clinical tag expansion using GPT-5.1, but the paper gives limited detail on prompt design, variability, or sensitivity of this component; and (5) while results are strong, the evaluation could be more convincing with statistical significance testing, per-class analysis, and stronger evidence that gains hold beyond this curated single-label setup.

  • 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.

  • 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 acceptance because the paper addresses a clear and important problem in rare retinal disease diagnosis, where standard retrieval often returns visually similar but pathologically mismatched references, and it proposes a well-motivated solution that directly targets this failure mode. The core idea of turning retrieval into a multi-step evidence refinement process via an MDP with DELETE/INSERT/TERMINATE actions is interesting and more appropriate than one-shot top-K retrieval for this setting. The graph-based state and homogeneity-aware reward are also well aligned with the goal of building a diagnostically coherent support set rather than relying only on visual similarity. Empirically, the gains over strong retinal foundation-model, static-retrieval, and PEFT baselines are substantial, and the ablations support that the key design choices matter. While I do have concerns about the lack of stronger clinical validation and the reliance on curated retrospective data, I think the paper’s problem formulation, technical design, and experimental improvements make it a meaningful and relevant contribution for MICCAI

  • Reviewer confidence

    Somewhat confident (2)

  • [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 AC and reviewers for recognizing Evo-RAD as a novel and meaningful step beyond static retrieval. All quantitative clarifications below are from original experimental logs, saved inference arrays, and analysis records omitted due to space limits. We did not change the model, data split, training protocol, or evaluation setting after review.

To Review 1:

Q1: About dataset. Retina-31 and Rare-20 are curated from the ASRS Retina Image Bank, a globally crowdsourced repository, rather than a single hospital. They inherently possess multi-center, multi-device, and diverse characteristics, thus representing distinct heterogeneous distributions.

Q2:Computational cost. Evo-RAD adds minimal overhead, increasing latency from 9.27ms to 15.86ms per image: a 6.59ms difference where each iterative update costs only 0.66 ms. Notably, the foundation model is run once for initial embeddings and is not re-executed during iterations. Evo-RAD maintains high efficiency at 3780 images per minute, fully satisfying the throughput demands of real-world clinical applications.

To Review 2:

Q1: Top-K retrieval followed by a trainable refinement. This simple baseline achieves only 56.57±0.37% on Retina-31, significantly trailing Evo-RAD’s 65.33±0.28%. This gap proves our performance gain stems from more than just a learnable scorer; while static re-rankers are restricted to a fixed, often noisy initial set, Evo-RAD dynamically optimizes evidence composition. By using INSERT to recover missing relevant data and DELETE to purge visually similar distractors before TERMINATE, it overcomes the inherent limitations of static sets where the correct sample is often initially missing or overwhelmed.

Q2: Quantitative metrics for initial retrieval. We evaluated the initial static retrieval and found a Recall@8 of only 44.57%. Because the diagnosis relies on hard voting, this low initial purity means correct samples are heavily overwhelmed by noisy distractors, confirming that fetching new candidates is essential rather than over-engineered. By the terminal step, Evo-RAD successfully increases the recall to 55.76±1.25%. This substantial +11.19% absolute gain demonstrates that our dynamic agent effectively purges misleading references and inserts coherent rare-disease evidence, far exceeding the capabilities of a static fixed-set re-ranker.

To Review 3:

Q1: Clinical Validation. Thanks for the suggestion. Prospective clinical trials and reader studies are beyond this paper’s current scope and reserved for future work.

Q2: Dependency on Database Quality & Noisy Labels. Evo-RAD mitigates dependency on database quality via its DELETE action, which functions as an intrinsic noise filter. Unlike passive static retrieval, Evo-RAD removes contradictory evidence, yielding a +3.9% improvement under 10% label noise. Crucially, this relative margin widened to +8.7% at 20% noise, validating its robustness.

Q3: Training Objective. Using ground-truth rewards is a common practice for RL training, yet the inference process remains fully autonomous without any label access.

Q4: Dependence on LLMs. LLMs are utilized only offline for descriptor generation and are not queried during inference, ensuring Evo-RAD is LLM-agnostic. Stability tests on Rare-20 confirm this, with consistent accuracy across Qwen-3 (45.98%±0.43%), GPT-5.1 (46.28%±0.56%), and GPT-5.5 (46.56%±0.45%).

Q5: Statistical Significance & Per-Class Analysis. Paired t-tests confirm statistically significant improvements over best-performing baselines, including RAC (p=0.0077) and static retrieval (p=0.0108), while per-class analysis highlights remarkable robustness in long-tail scenarios, specifically on rare diseases like Retinoschisis (+30.71% ACC, only n=16 training images), Macular dystrophy (+24.89%, n=14), Choroideremia (+16.32%, n=14), and Susac’s syndrome (+14.20%, n=15), proving iterative evolution effectively circumvents clinical data scarcity.




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 paper addresses an important rare retinal disease diagnosis problem and proposes a clearly motivated agentic retrieval framework that reformulates evidence selection as a sequential refinement process. The reviewers generally recognize the novelty of the MDP formulation, the graph-based policy, and the homogeneity-aware reward, and the reported gains over foundation-model, static-retrieval, and PEFT baselines are promising. The main unresolved issue is whether the full multi-step agent is necessary compared with simpler alternatives, such as Top-K retrieval followed by a trainable re-ranking/refinement module. The authors should also report initial Recall@K/Hits@K, computational cost, latency, statistical significance, and more detailed per-class or robustness analyses. If the rebuttal can clarify these points, especially the necessity of INSERT/DELETE/TERMINATE actions beyond simpler re-ranking, the paper would be suitable for acceptance.

  • 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 directly addresses the main concern by adding the missing re-ranking baseline and initial retrieval metrics, showing that the full iterative INSERT/DELETE/TERMINATE design improves beyond static refinement. Given the strong post-rebuttal support from reviewers and the clarified efficiency, robustness, and statistical significance, the paper is suitable for acceptance.



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.

    Following the rebuttal round, the paper now has three accepts. I also 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.

    All reviewers provided final positive scores.



back to top