Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Generalized Category Discovery (GCD) aims to identify known disease categories while discovering unseen ones in medical images, without relying on exhaustive manual annotations. However, existing GCD methods for medical imaging are typically evaluated under balanced settings, overlooking a critical characteristic of practical scenarios: novel categories often correspond to rare diseases and exhibit severe long-tailed distributions. To address this limitation, we propose RareGCD, a framework tailored for long-tailed medical data. Specifically, RareGCD leverages labeled data from known categories as a strong prior to guide sample association, enabling reliable grouping of unlabeled instances under explicit category constraints. Furthermore, we propose a prototypical contrast learning strategy with density-aware instance-wise temperature scaling, which strengthens contrast supervision for reliable samples and reduces over-confident updates for boundary samples, improving representation stability under imbalance. To better reflect practical scenarios, we construct long-tailed benchmarks on multiple medical datasets and evaluate RareGCD under both standard and long-tailed settings. Extensive experiments demonstrate that RareGCD consistently outperforms state-of-the-art methods, particularly in recognizing novel disease categories.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1264_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/myyy777/RareGCD

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MaYua_RareGCD_MICCAI2026,
        author = { Ma, Yuan AND Feng, Wei AND Hou, Junlin AND Zhang, Chao AND Ge, Zongyuan AND Xie, Haoran AND Ju, Lie},
        title = { { RareGCD: Toward Rare Disease Discovery via Generalized Category Discovery } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This work tackles the problem of class imbalance in Generalized Category Discovery (GCD) and proposes RareGCD. The method enhances prototype learning with an instance-aware temperature mechanism to address the issue of imbalanced distributions. Experiments on the PathMNIST, OrganAMNIST, and BloodMNIST datasets demonstrate that RareGCD achieves competitive performance under both standard and long-tail settings.

  • 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 method achieves strong results on three publicly available datasets.
  • 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.

    • Limited novelty. 1.Some prior works have addressed class imbalance in GCD (e. g. , [R1, R2]). The authors should better emphasize the theoretical differences and experimental advantages of RareGCD over existing methods. 2.Prototype learning, two-stage training, and density-aware mechanisms are commonly used in previous studies. The authors should clearly summarize the contributions of this work in the Introduction. [R1] ImbaGCD: Imbalanced Generalized Category Discovery [R2] Long-Tailed Learning for Generalized Category Discovery

    • Method design issues: Section 2.4 employs entropy maximization for Decision-Boundary Refinement. However, entropy maximization mathematically assumes uniform class distributions, which conflicts with the class-imbalance problem that the method aims to solve.

    • Parameters such as $\beta$ and $\lambda$ are set to 0.1 and 0.35, respectively, but there is no ablation study or explanation of their impact on results.

    • Concerns on experimental design: The three datasets do not include real rare diseases, and inter-class differences are pronounced. The current evaluation simulates rare classes solely via data imbalance, whereas true rare diseases often closely resemble common ones, making them harder to distinguish. Therefore, it is recommended to describe the method as addressing class imbalance rather than emphasizing rare diseases.

  • 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?

    Considering the limited novelty and the methodological aspects that require further clarification, my overall assessment is weak reject.

  • 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 author’s response addressed most of my concerns, and I agree to accept this paper.



Review #2

  • Please describe the contribution of the paper

    The paper proposes RareGCD a generalised category discovery (GCD) framework for identifying novel categories corresponding to rare diseases under long-tailed distributions. The paper introduces prototype-guided sample association, a density-aware prototypical contrastive learning objective to stabilise representation learning under severe class imbalance, and a decision-boundary refinement via a parametric classifier with self-distillation. The paper conducts controlled experiments on 3 datasets, demonstrating consistent performance improvements over SOTA GCD methods, particularly on novel and rare categories.

  • 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 is well written and technically sound. Well-motivated methodological design addressing long-tailed distributions in medical settings where novel categories correspond to rare diseases.

    • The instance-wise temperature scaling based on local feature density is meaningful and intuitive.

    • The method is evaluated across three datasets PathMNIST, OrganAMNIST, and BloodMNIST in standard and long-tailed settings. The proposed method RareGCD outperforms over baselines. The ablation study support the importance of each proposed component.

  • 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.

    • Limited novelty relative to prior GCD frameworks. The novelty lies more in combination and adaptation components build incrementally on existing ideas such as prototype-based association, reciprocal KNN, and parametric classification for GCD [1,2,3] rather than fundamentally new theory.

    • Hyperparameter sensitivity not fully analyzed. the temperature bounds (τ_min, τ_max), association thresholds λ=0.35, and fixed weights β=0.1 set to fixed without proper justification.

    • The EMA-based prototype memory E update (Sec. 2.2, page 5) may still reinforce early association errors and bias representations toward head classes under severe long-tailed distributions, as it relies on point-estimate prototypes without explicit uncertainty or reliability modeling for rare categories.

    • Similar to prior GCD work, the experiments assume that the total number of categories (using KNN) is known, which limits applicability in realistic clinical discovery scenarios.

    • Minor weakness: The table and figure captions require more clarity. The controlled experimental settings should be explicitly highlighted in the abstract, introduction, and/or conclusion.

    [1] Vaze, S., Han, K., Vedaldi, A., Zisserman, A.: Generalized category discovery. CVPR (2022) [2] Wen, X., Zhao, B., Qi, X.: Parametric classification for generalized category discovery: A baseline study. CVPR (2023) [3] Wang, H., Vaze, S., Han, K.: Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. arXiv preprint arXiv:2403.13684 (2024)

  • 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?

    N/A

  • 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 clarified their contribution and differences from prior work in the rebuttal. I recommend adding the ‘parameter ablation study’ to the main paper to improve reproducibility.



Review #3

  • Please describe the contribution of the paper

    This paper extends medical Generalized Category Discovery to a more realistic long-tailed distribution setting and proposes RareGCD to address the resulting challenges. The core and sole methodological contribution lies in Density-Aware Prototypical Contrast Learning, together with a two-stage training paradigm. Extensive experiments demonstrate the effectiveness of RareGCD.

  • 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 problem addressed by RareGCD is well-motivated and carries clear clinical significance.

    2.The proposed method is sound and easy to follow. The Sample Association strategy is effective in reliably identifying correct pairings.

    3.Although the approach involves a relatively large number of hyperparameters, the overall design is well-justified.

  • 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.Some key details necessary for reproducibility need to be clarified. Such as, the criteria for determining “high neighborhood agreement”, “lower reliability score” are not explicitly defined. In addition, many hyperparameters are insufficiently described, along with the principles for their selection—for example, the threshold parameter (thresh) in the code and the weighting coefficient in the density-aware (DA) loss.

    2.On Page 5, Line 4, “μ E” should be written as “μ E [c_i]”. On Page 4, Line 22, ( c_i \in C ) is incorrect since ( C ) is presented as a scalar rather than a set. Furthermore, there are numerous inconsistencies in terminology, such as “Sim-GCD” vs. “SimGCD”, “SPTNet” vs. “SPTNET”, and “MedGCD” vs. “Med-GCD”.

    3.A critical concern is that, under a long-tailed distribution, rare diseases with very few samples are treated as novel categories; accurately estimating the number of categories in this setting is an essential step. The paper notes that MedGCD(2025) and GCD(2022) assume the number of categories is known a priori; however, these works do not operate under long-tailed distributions, and GCD(2022) in fact proposes a method for estimating the number of categories. This discrepancy is not adequately addressed.

    4.Moreover, the method introduces a large number of hyperparameters without providing reliable selection strategies, raising the possibility that the reported performance gains are partly attributable to extensive tuning.

    5.From a methodological perspective, the approach is largely incremental and lacks substantial theoretical novelty. The two-stage training procedure also introduces additional computational overhead.

    6.The paper would benefit from more comprehensive visualizations to support its claims.

  • 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 paper is well-written, and the proposed method is sound. The problem it addresses is practically relevant. Despite the lack of many key implementation details, it still offers valuable insights to the research community. Therefore, I maintain a positive overall assessment.

  • 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.

    Thank you for the authors’ response. I believe the rebuttal has substantially addressed my concerns, and the task introduced by RareGCD carries considerable clinical significance. However, the authors should more clearly specify every previously omitted detail, as this is crucial for readers. In addition, since this is presented as a new task, the experimental evaluation should be more comprehensive. Estimating the number of categories in GCD is a widely studied and important experiment rather than a simple ablation, and it is particularly critical in the long-tailed distribution scenario of rare medical diseases.



Author Feedback

We thank the Reviewers R1, R2, and R3 for their valuable feedback. R1 & 2 & 3: Q1: Methodology. A: Thanks for the valuable points. We would like to clarify that our contribution is not a simple assembly of existing components, but three designs specific to long-tailed medical GCD: (i) a constraint-preserving prototype-guided association that prevents minority classes from being absorbed by majority ones, (ii) a density-aware instance-wise temperature for prototypical contrast that adapts to per-sample reliability, (iii) the first long-tailed GCD benchmark on medical datasets, on which RareGCD sets the new SOTA. The two stages share the same backbone and total epochs as the baselines, resulting in acceptable extra training and inference cost. Q2: Parameter ablation. A: The parameters are set as follows: β = 0.1, λ = 0.35, τ_min = 0.1, τ_max = 0.7, memory update rate μ = 0.2, threshold is 0.7 for all experiments. We conducted parameter ablation before submission but omitted them from the paper due to the page limit. Using PathMNIST (r=100) as an example, the overall accuracy fluctuated within 2% when β ranged from 0.1 to 0.3, within 2.5% when the threshold ranged from 0.65 to 0.8, and within 1.8% when τ_min and τ_max ranged from 0.1–0.2 and 0.7–0.8, respectively. These results indicate that RareGCD is not sensitive to these parameters. Our code has been released to support reproducibility. R1: Q3: Differences from prior works. A: ImbaGCD performed class-prior alignment with optimal transport to correct biased cluster assignments, while LT-GCD introduced self-guided labeling and representation balancing. However, both methods implicitly rely on the learned cluster structure. Under severe long-tailed settings, this assumption can be fragile, as tail classes may be absorbed by head classes in training. RareGCD addresses this issue through constraint-preserving association and density-aware learning. We did not conduct comparison with these works because their codes were not publicly available. Q4: Entropy maximization. A: The entropy maximization term is not responsible for handling class imbalance; it serves as a weak anti-collapse regularizer in the decision-boundary refinement stage. The long-tailed issue is addressed by the prototype-guided association and density-aware prototypical contrast learning. Q5: Rare diseases simulation design. A: We agree that the current datasets simulate rarity via long-tailed class distributions rather than representing clinically confirmed rare diseases. We will revise the claim to “rare-category discovery under long-tailed medical data” and clarify that our experiments provide a controlled proxy for rare-disease-related data scarcity. R2: Q6: EMA update. A: Directly adding uncertainty-weighting to EMA tends to slow down rare-class prototype formation, which already suffers from limited samples. RareGCD instead uses density-aware instance-wise temperature to reduce noisy supervision without suppressing rare-class prototype formation. Q7: Table captions. A: We will refine the captions and highlight the long-tailed setting. R2 & 3: Q8: Estimating the number of classes. A: The citation in our paper will be revised to reference works that explicitly adopt the known-K assumption (e.g., SimGCD, MedGCD). In previous experiments, the K-estimation results of GCD (2022) are unreliable under severe imbalance. Specifically, on PathMNIST, OrganAMNIST, and BloodMNIST (r=100), the estimation errors are 33%, 27%, and 37%, respectively. We will leave long-tailed K-estimation as future work. R3: Q9: Key details. A: High neighborhood agreement refers to candidate pairs whose Jaccard distance is below a threshold. Reliability score is a ratio of intra-cluster similarity mass. Q10: Writing inconsistencies. A: We will correct the equations on Pages 4 and 5, and unify the method names as SimGCD, SPTNet, and MedGCD. Q11: Visualizations. A: We will include visualizations of learned feature embeddings.



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 your paper’s clinical relevance and find the density-aware methodology to be a sound approach that achieves strong performance across multiple datasets. To strengthen your submission during the rebuttal without needing to generate new experimental results, we gently suggest focusing on clarifying the theoretical novelty of your work compared to prior generalized category discovery frameworks. Specifically, please provide a clearer theoretical justification for your hyperparameter choices and explain the rationale behind using entropy maximization in a class-imbalanced setting. Additionally, addressing the reviewers’ concerns regarding the assumption of a known number of categories , refining the terminology around “rare diseases” versus simulated data imbalance , clarifying missing implementation details such as the EMA-based prototype updates , and correcting the noted typographical errors will comprehensively resolve their core questions.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    While the reviewers recognized the clinical relevance of your application and appreciated the strong empirical performance achieved by your density-aware methodology across multiple datasets, the final consensus is that the submission falls short of the required threshold for publication. The primary reason for this decision is the lack of substantial theoretical and methodological novelty, combined with several critical design contradictions and unresolved experimental concerns. Primary Weaknesses Limited Technical Novelty (Incremental Contribution) The technical components of the proposed framework build incrementally on top of existing rameworks. The novelty lies more in the combination and adaptation of existing modules rather than a fundamentally new theory. The submission fails to clearly emphasize the theoretical differences and explicit methodological advantages of RareGCD over existing frameworks that already address class imbalance in GCD.



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 proposes RareGCD, a generalized category discovery framework for rare-category discovery in long-tailed medical image data. The method combines prototype-guided sample association, density-aware prototypical contrastive learning, and decision-boundary refinement, and reports consistent improvements on three medical image datasets.

    The initial reviews were mixed but overall close to positive, with scores of 3 / 4 / 4.Reviewers appreciated the clinical relevance, clear motivation, and empirical performance, but raised concerns about limited novelty over prior GCD methods, insufficient hyperparameter analysis, missing implementation details, the use of simulated long-tailed settings rather than real rare-disease cohorts, and the assumption/estimation of the total number of categories.

    The rebuttal addressed several key concerns. The authors clarified the methodological contribution and differences from prior work, explained the role of the entropy term, provided parameter sensitivity results, and agreed to soften the rare-disease claim toward rare-category discovery under long-tailed medical data. These responses were viewed positively by the reviewers.

    Some limitations remain. In particular, category-number estimation under long-tailed settings is still not fully evaluated and is left for future work. The evaluation remains a proxy for rare disease discovery, and the camera-ready version should include the promised parameter sensitivity results, implementation details, and clearer discussion of limitations.

    After rebuttal, all three reviewers recommend Accept, with final scores of A / A / A. Reviewer 1 moved from Weak Reject to Accept, mainly due to the clarification of contributions and responses to prior concerns. Reviewers 2 and 3 also maintained positive assessments, while noting that reproducibility details and the category-number estimation issue should be clarified.

    I recommend acceptance. The reviewer consensus after rebuttal is clearly positive, and the remaining issues are limitations that can be addressed in the final version rather than reasons for rejection. The paper addresses a relevant and timely problem, proposes a sound framework, and provides consistent empirical evidence, although the authors should temper the rare-disease framing and clearly state the remaining limitations.



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 reviewers have reached a consensus to accept the paper.



back to top