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

Abstract

Medical ordinal classification aims to predict ordered stages of disease progression. Existing ordinal metric learning methods enforce globally monotonic embeddings with fixed inter-stage margins to preserve stage ordering, implicitly assuming uniform stage transitions. However, real-world medical datasets are typically long-tailed, making severe stages rare yet clinically critical. Applying fixed ordinal constraints leads such imbalanced data to structural incompatibility: fixed margins cause boundary compression for sparse severe stage transitions, while data imbalance biases rank embeddings against rare stages. To address these issues, we propose the Frequency-aware Multi-Margin Loss (FMML). FMML introduces Boundary-Level Scaling to adapt inter-stage margins based on transition sparsity and Anchor-Level Scaling to strengthen supervision for rare stage anchors, effectively preserving ordinal consistency with long-tailed adaptation. Furthermore, since metrics like MAE fail to capture the systematic directinoal bias, we introduce Long-Tailed Rank Deviation (LTRD). This metric evaluates ordinal classification tasks while accounting for both rank errors and data distribution. Experiments on four imbalanced medical grading datasets demonstrate that the proposed method achieves superior performance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4760_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yinpub/ltrank2026

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YinHao_FrequencyAware_MICCAI2026,
        author = { Yin, Haojie AND Liu, Tianyi AND Feng, Chengcheng AND Huang, Kaizhu},
        title = { { Frequency-Aware Multi-Margin Loss for Imbalanced Medical Ordinal Classification } },
        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 addresses the challenge of ordinal classification in imbalanced medical datasets by proposing a Frequency-Aware Multi-Interval Loss (FMML). By integrating boundary-level and anchor-level scaling mechanisms, the method adaptively adjusts geometric intervals between disease levels, effectively mitigating the compression of minority class feature boundaries often found in traditional metric learning. To better capture the directional biases of minority classes—which standard metrics like MAE fail to reflect—the authors further introduce a specialized evaluation metric, Long-Tail Rank Deviation (LTRD). Systematic evaluations across four medical datasets (GAMMA, BUSI, APTOS 2019, and KOA) demonstrate that FMML consistently outperforms existing mainstream methods in terms of accuracy, MAE, and F1 scores, with ablation studies confirming the synergy of its dual-scaling components.

  • 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 article clearly points out the inherent conflict between “ordinal constraint” and “long-tail correction”, which is often overlooked in existing work. Especially in medical classification tasks, although severe cases are rare, the cost of misdiagnosis is extremely high. The author poses the problem from this practical need, with sufficient motivation and practical significance. 2.The method design exhibits strong pertinence and innovation. The author introduces the frequency perception mechanism to both the geometric level (Boundary-Level) and the optimization level (Anchor-Level), which complement each other. This approach not only adjusts the inter-class spacing but also strengthens supervision during the training process, effectively avoiding the shortcomings of previous methods that only addressed the long-tail problem at the classifier level but were disconnected from embedding space learning. 3.In terms of evaluation metrics, the proposed LTRD is somewhat enlightening. Compared to MAE, which only focuses on absolute error, LTRD further considers the directionality of prediction bias and mitigates the dominant role of head categories through frequency weighting, making it more suitable for evaluating long-tail ordinal tasks. It can be seen as a beneficial supplement to the existing metric system. 4.The experimental section is solid and persuasive overall. The author conducted validations on four different types of medical datasets, and the results consistently outperformed multiple recent methods. Meanwhile, through ablation studies and parameter sensitivity analysis, the author fully demonstrated the role of each module and their synergistic effects. Combined with various metrics such as Acc, MAE, LTRD, and F1 scores for different categories, the author presented the model’s performance from different perspectives, balancing stability in head categories and improving performance in tail categories.

  • 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.There is still room for further improvement in baseline selection and comparison settings. Although the article covers various recent methods, it is not clear whether works like CLOC and PCOL from 2025 adopted their respective optimal implementation versions; meanwhile, it also requires clearer explanations on whether the same backbone network (such as ResNet-50) and consistent data augmentation strategies are uniformly used across different methods, otherwise it may affect the fairness of the comparison. 2.The discussion on clinical value and model generalization ability is relatively insufficient. The article repeatedly emphasizes the “clinical criticality” of the method, but lacks interpretation of results from clinical experts or alignment analysis with actual clinical scoring systems (such as ophthalmological grading standards). Furthermore, the experiment is solely based on ResNet-50, without further exploration of stronger models (such as ViT) or methods incorporating self-supervised pre-training, making it difficult to determine whether the method can further amplify its advantages under a higher-performance framework. 3.The description could be more specific regarding implementation details and hyperparameter settings. For instance, the selection and tuning process of key parameters such as λ in LTRD, weight coefficients in the overall loss, as well as δ and γ, are not elaborated, which to some extent affects the reproducibility of the method. 4.Lack of discussions or comparisons with previous ordinal regression methods like [1], [2], [3], etc.

    [1] CORE: Learning consistent ordinal representations with convex optimization for image ordinal estimation, Pattern Recognition, 2024. [2] Soft labels for ordinal regression, in: Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , 2019. [3] Learning probabilistic ordinal embeddings for uncertainty-aware regression, CVPR 2021.

  • 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.

    (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?

    “Imbalance” and “Ordinal relationships” are critical to medical area, and the proposed Long-Tail Rank Deviation is novel to measure this issue quantitatively.

  • 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 #2

  • Please describe the contribution of the paper

    The paper addresses ordinal classification under long-tailed distributions in medical image grading. The authors identify structural incompatibility between ordinal emtric learning and standard long-tail corrections.

    The proposed FMML resolve this using learnable margin and boundary-level scaling and anchor-level scaling.

    The paper also propose LTRD a frequency-aware metric that capture under-estimation of rare stages.

    The paper have consistent performacne across four imbalanced medical 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.
    • Insightful and clinically grounded motivation
    • Clearly identify the structral-conflict between ordinal grade and long-tail correction (Figure 1). It is an interesting read.
    • Nice separation of the proposed scaling; the ablation support their complementarity
    • LTRD makes sense and can be used in the future
  • 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 don’t think there is “adaptiveness” in the scaling as it is computed from trianing-set class counts and remain fix
    • FMML relies on access to training-set class-frequency statistics, while most baselines do not exploit this prior to the same degree; I wonder whether the gains stem from the novel formulation or simply a stronger prior
    • The learnable margins might collapse to <= 0, there is no mechanism described to prevent that
    • not sure how robust the method is to the train/test distribution mismatch, especially considering the prior of scaling factor
    • the novelty is moderate as individual components have been proposed else where, e.g. inverse-frequency scaling and learnable multi-margin ordinal losses
  • 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.

    (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?

    In sum, this is a clinically well-motivated paper with a clear thesis and consistent empirical support. The structural-conflict framing and the cumulative-axis construction are good contributions, and LTRD is a useful metrics. Althought the technical novelty is moderate, the clinical grounding and the breadth of datasets makes it a great fit for MICCAI.

    In the final version or rebuttal, I would like to see discussion more on

    1.margin positivity 2.adaptiveness of the scaling itself 3.generalisability under distribution shift; note that in real calinical settings prevalence varies by site, population etc, so, test-time class proportions often differ substantially from training. I would like to learn more about the impact of such mismatch 4.the ablation in table 3 only remove scaling; would it be possible to isolate the contribution of multi-margin?

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The author has answered my questions and I am happy to champion this paper.



Review #3

  • Please describe the contribution of the paper

    This paper studies imbalanced medical ordinal classification and proposes Frequency-Aware Multi-Margin Loss (FMML), which augments ordinal metric learning with two frequency-aware components: Boundary-Level Scaling to enlarge sparse adjacent-stage boundaries, and Anchor-Level Scaling to strengthen rare-class anchors during optimization. The paper also introduces Long-Tail Rank Deviation (LTRD) as a metric intended to capture directional bias and class-frequency effects beyond MAE. Experiments are conducted on four medical grading datasets: BUSI, GAMMA, APTOS 2019, and KOA.

  • 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 method is simple and reasonably well aligned with the problem. The decomposition into optimization-side correction (anchor scaling) and geometry-side correction (boundary scaling) is intuitive. In particular, the idea of modulating adjacent ordinal margins according to local boundary density is a natural way to preserve ordered structure while adapting to imbalance. This is one of the stronger aspects of the paper.

    2) The paper attempts to go beyond MAE with LTRD. The motivation for LTRD is sensible: MAE cannot distinguish unbiased errors from consistent drift toward dominant stages. A metric that explicitly captures directional bias and class-frequency balancing is potentially useful for long-tailed ordinal medical tasks.

    3) The paper conducts evaluations across four public datasets with varying degrees of class imbalance, which helps support the robustness and consistency of the proposed method. The inclusion of multiple evaluation metrics (Accuracy, MAE, F1-score, and the proposed LTRD) provides a relatively comprehensive view of performance, particularly in capturing both overall accuracy and ordinal-specific errors.

  • 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 novelty over CLOC (CVPR 2025) appears limited. CLOC has already established the core paradigm of multi-margin contrastive learning for ordinal classification, explicitly modeling ordered class structure in the embedding space. In comparison, the proposed FMML mainly introduces frequency-aware scaling on margins and anchors, together with the new LTRD metric. While these extensions are relevant to imbalanced medical settings, they seem more like a task-specific refinement of the existing CLOC framework than a fundamentally new ordinal learning approach.

    2) LTRD is one of the advertised contributions, but the validation is still weak. The paper defines it as a weighted combination of class-wise rank error plus directional bias penalty, reweighted by inverse square-root frequency. However, the newly proposed metric (LTRD) requires more careful and rigorous justification. In particular, it would be important to compare it against widely adopted ordinal evaluation metrics such as Quadratic Weighted Kappa (QWK), which is commonly used in medical grading tasks (e.g., diabetic retinopathy assessment).

    3) The introduction emphasizes the importance of effectively handling minority (tail) classes under long-tailed distributions. However, this claim is not sufficiently substantiated in the experimental section. In particular, the paper does not clearly demonstrate whether the proposed method provides improvements over existing approaches in recognizing rare classes. To better support this core motivation, it would be highly beneficial to include additional experiments specifically focused on tail-class performance. For example, reporting per-class precision and recall, especially for the rarest categories, would provide more direct evidence of improvement. Furthermore, metrics such as macro-averaged F1-score or tail-class averaged performance could help highlight gains that may be obscured by overall accuracy. Such analyses would make the contribution more convincing and better aligned with the problem statement.

  • 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.

    (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 refer to the previously discussed strengths and 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

1.Novelty @R2, R3 FMML is novel in reframing long-tailed (LT) ordinal task from class-level reweighting to boundary-aware frequency correction. Although reweighting and margin-based ordinal losses have been studied, directly combining them is ineffective because they act on different units: LT losses address class-frequency bias at the class level, while ordinal losses focus on transitions between adjacent stages. Such naive combinations may overemphasize rare stages as isolated classes, improving separation but distorting the ordinal structure. FMML resolves this mismatch by combining frequency-aware correction with ordinal boundary geometry. Boundary-level scaling widens sparse adjacent intervals, and anchor-level scaling strengthens rare-stage supervision while preserving ordinal relations. 2-3 @R3 2.Role of LTRD QWK and MAE measure global ordinal errors but miss clinically important asymmetric errors in LT medical tasks. E.g., QWK penalizes misclassifying a normal case (head class) as late stage (tail class) and a late stage case as normal equally, although the latter is more harmful because it underestimates severity. LTRD addresses this by weighting tail-class distance errors. 3.Tail-class performance We respectfully clarify tail-class evidence has been already included in the submission. In Sec. 3, mostly results are reported using per-class F1-scores, including tail classes. 4-5 @R1 4.New experiments results We add new experiments on Aptos, and other datasets show similar trends. We will report full results in revision. 1) Earlier Ordinal Methods FMML still outperforms earlier methods (ACC/MAE/LTRD): Ours 0.8579/0.1913/0.8509 vs. POE 0.8306/0.2218/0.9025, CORE 0.8388/0.2049/0.8792, and SORD 0.8279/0.2186/0.9319, indicating gains still beyond earlier baselines. 2) Backbone Generality We show FMML remains effective with ViT backbone (vit_small_patch16_224). Under same protocol, FMML achieves best ACC/MAE/LTRD (0.7404/0.4317/1.7179) vs. CLOC 0.7295/0.4590/2.0310, ORL 0.7022/0.4863/2.1610, and PCOL 0.6092/0.6202/2.5739.While ViT’s absolute performance is lower than ResNet (due to extreme imbalance thus causing overfitting), the consistent relative gain supports backbone generality. 5.Baseline Setting We clarify the comparison is fair: all methods use the same backbone & augmentation. We set lambda to 0.8 to balance the directional bias and rank error in LTRD. A larger lambda increases sensitivity to systematic over- or under-estimation, while a smaller one makes LTRD closer to class-wise rank error. In our setting, a higher lambda (e.g. 0.6~1) is appropriate when error direction is clinically more important. Also, we clarify that all the conclusions are consistent when 0.6<lambda<1.6-9 @ R2 6.Relationship between gains & prior Many baselines (e.g., BAW, WCE, LOS, and POCL) also leverage class-frequency prior. This observation suggests our gains are attributable to the novel design, rather than to the use of priors. We will add more detailed description of baselines. 7.Adaptive Scaling Sorry for confusion. We use adaptive scaling factors to denote that different samples are modulated by different scaling factors based on class frequency. The scaling strength for each class is computed from the data distribution and kept fixed during training. The margin is adaptive in that different classes are assigned different margins. 8.Margin Positivity The margins cannot be negative. The raw margin are re-parameterized by softplus. (see line 89 in our code loss/fmmoc.py) 9.Distribution shift During outbreaks of emerging diseases, training data may be extremely LT, while the prevalence of severe cases may increase after deployment, leading to a train–test distribution shift. This shift can make inference harder, e.g., when patients labeled early during training later progress to severe stages. Such evolving shifts remain underexplored. We will discuss them in revision.



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.

    This paper presents a method for imbalanced medical ordinal classification. While all reviewers commented positively on the method contribution, there are several important questions to be addressed. For example, there should be more discussion in comparison with other related methods and method novelty should be further elaborated. Validation of claims should be stronger, particularly regarding LTRD and effective handling of tail classes.

  • 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.

    After the rebuttal, the reviewers’ ratings have not changed. While the negative reviewer did not provide a final rating after rebuttal, the authors’ rebuttal appears to have answered the reviewer’s comments. Considering that all reviewers recognize the method contributions and extent of evaluation on four datasets, it’s recommended to accept the paper.



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.

    I recommend accept: R2 maintained Accept and confirmed that the closest prior work is different from this paper’s new ideas—directly addressing the central novelty concern—while R1 and R3 did not update.



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.

    After considering the reviews and the rebuttal, I recommend acceptance. The paper presents a clinically motivated contribution for imbalanced medical ordinal classification, with useful ideas on boundary-aware frequency correction and the LTRD metric. The rebuttal clarifies the novelty relative to related ordinal methods, the role of LTRD, tail-class evidence, baseline settings, and margin positivity. While the novelty is not ground-breaking, the work is well motivated, empirically supported across multiple datasets, and represents a meaningful contribution to the field.



back to top