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

Abstract

Multi-label classification (MLC) is a central problem in medical image analysis (MIA), as a single medical image often contains multiple coexisting abnormalities. While ranking-based loss methods have been widely adopted for MLC, their effectiveness in MIA is limited by two major challenges: the dominance of negative gradients during optimization, and the uniform suppression of all negative logits without accounting for class-specific logit distribution. To address these limitations, we propose Reference-Guided Gradient Balancing (RGB) loss, a novel loss function designed for long-tailed multi-label medical image classification, which comprises an adaptive negative suppression loss and a positive enhancement loss. The adaptive negative suppression loss operates based on whether the correction and incorrect prediction coexist within a batch. When both are present, correct predictions serve as references that reveal the class-specific logit distributions and guide the rectification of hard sample logits, thereby refining the decision boundary of each class. Otherwise, negative gradients are adaptively reduced based on prediction confidence. Additionally, the positive enhancement loss incorporates sample difficulty and class frequency to enhance the recognition of low-prevalence diseases. Extensive experiments on NIH ChestXRay14, CXR-LT, and CheXpert demonstrate that RGB consistently outperforms state-of-the-art methods, with further analysis confirming the effectiveness in suppressing negative gradients and increasing the separation between positive and negative logits. The code is available at https://github.com/xup6YJ/RGB-Loss.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0566_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/xup6YJ/RGB-Loss

Link to the Dataset(s)

CXR-14: https://openaccess.thecvf.com/content_cvpr_2017/papers/Wang_ChestX-ray8_Hospital-Scale_Chest_CVPR_2017_paper.pdf CXR-LT: https://www.sciencedirect.com/science/article/pii/S1361841525002865 CheXpert: https://stanfordmlgroup.github.io/competitions/chexpert/

BibTex

@InProceedings{LinYin_ReferenceGuided_MICCAI2026,
        author = { Lin, Ying-Chih AND Peng, Hsiao-Chu AND Kuo, Po-Chih AND Chen, Yong-Sheng},
        title = { { Reference-Guided Gradient Balancing Loss for Long-Tailed Multi-label Medical Image Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a novel Reference-Guided Gradient Balancing loss for imbalanced MLC in MIA. The proposed RGB loss consists of the following main contributions:

    1. Positive Enhancement Loss which replaces the standard class-level separation term with a confidence-weighted ranking loss, and dynamically adjusted by class frequency to amplify gradient signal for rare classes relative to common ones. 2.Adaptive Negative Suppression Loss which uses correct predictions (TP or TN) as reference signals. False positive logits are pushed below TP logits when TPs exist, or below TN logits otherwise. Uses correct predictions (TP or TN) as reference signals. False positive logits are pushed below TP logits when TPs exist, or below TN logits otherwise. This provides class-specific, data-driven suppression thresholds. 3.Consistent SoTA on all three CXR benchmarks across all class groups, with the largest gains on tail classes.
  • 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 is well motivated and relatively easy to follow. 2.Leveraging TP/TN predictions within the same batch as dynamic, class-specific anchors is an elegant idea. 3.RGB achieves SoTA on all three benchmarks across different class groups. 4.RGB is a pure loss function replacement, so it should be easy to be used, which is a significant advantage for clinical adoption.

  • 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 paper frames its problem as long-tailed multi-label classification, but this characterization is questionable for medical imaging datasets. Long-tailed learning literature [1] typically targets datasets with a very large number of categories where the label frequency follows a heavy-tailed distribution. Medical imaging datasets, however, contain a limited number of disease labels , which is more accurately characterized as a class imbalance problem [2]. The distinction matters because long-tailed methods make implicit assumptions about label cardinality that do not hold here. The authors should either justify the long-tailed framing more rigorously or reposition the problem as class-imbalanced multi-label classification. 2.The paper partitions classes into head/medium/tail groups using dataset-specific prevalence thresholds (10%/1.5% for CXR-14, 10%/2% for CXR-LT, 40%/5% for CheXpert) with no principled justification for these choices. It is unclear whether these thresholds were selected based on data statistics, prior work conventions, or tuned to favor the reported results. 3.Figures should be provided in lossless format. 4.The dynamic weighting parameter $\lambda = 1 - \frac{1}{\log_2(D_k)}$ is undefined when $D_k = 1$ ($\log_2(1) = 0$) and becomes negative when $D_k < 2$, which inverts the exponent sign and reverses the ranking objective. Crucially, this pathological case arises precisely for the rarest disease classes. 5.The paper describes $\lambda$ as a “dynamic gradient weighting hyperparameter,” but $D_k$ is a fixed dataset-level statistic computed before training. $\lambda$ is therefore a static, class-specific constant, not a dynamic quantity. 6.The prediction probability $p$ is defined as $\frac{1}{1+e^{-s_k}}$ using only the class index $k$, yet it is used as $p_i$ in Eq. 2 as if it were sample-dependent. The intended definition should be $p_{i,k} = \frac{1}{1+e^{-s_{i,k}}}$. 7.In the $\neg$TP branch of $\mathcal{L}{\text{ref}}$ (Eq. 4), the term $e^{s{j,k} - s_{l,k}}$ produces a gradient on TN logit $s_{l,k}$ that pushes it upward (relative to FP logits). Simultaneously, the second case of $\mathcal{L}{\text{sup}}$ (Eq. 5) applies $\log(1 + e^{s{l,k}})$ to all TN samples, pushing their logits downward. These two objectives are in direct conflict whenever FP predictions exist but no TP reference is available, which could cause gradient instability or cancel useful learning signal.

    1.Zhang, Y. , Kang, B. , Hooi, B. , Yan, S. and Feng, J. , 2023.Deep long-tailed learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 45(9), pp. 10795-10816. 2.Cong, C. , Liu, S. , Rana, P. , Pagnucco, M. , Di Ieva, A. , Berkovsky, S. and Song, Y. , 2024.Adaptive unified contrastive learning with graph-based feature aggregator for imbalanced medical image classification. Expert Systems with Applications, 251, p. 123783.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I think the paper is overall well-written and easy to follow. Besides, the proposed idea should be easy to be implemented and highlights potential advantage for clinical adoption. However, there are couple of unclear statements in the method section, the authors are encouraged to clarify these points during rebuttal.

  • 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.

    I would like to thank the authors for their effort in preparing the response. Most of my concerns have been adequately addressed.



Review #2

  • Please describe the contribution of the paper

    This paper proposes Reference-Guided Gradient Balancing (RGB) loss for long-tailed multi-label medical image classification. The main idea is to address two claimed weaknesses of prior ranking-based losses: dominant negative gradients in highly sparse multi-label settings, and uniform suppression of negative logits without accounting for class-specific logit distributions. RGB combines a positive enhancement term with an adaptive negative suppression term that uses correct predictions as references to rectify false positives. Experiments are conducted on ChestXray14, CXR-LT, and CheXpert using DenseNet121, and the paper reports improvements over several prior loss-function 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 paper addresses an important and practical problem. Long-tailed multi-label chest X-ray classification is clinically relevant, and the observation that gradients from abundant negative labels can overwhelm learning for rare diseases is well motivated. The proposed loss is also easy to understand and fairly well structured: the positive enhancement term directly targets rare and hard positives, while the adaptive negative suppression term uses true positives or true negatives as reference anchors to guide correction of false-positive logits. The ablation studies are helpful and show that both components contribute. The TP-or-TN reference design also appears better than using only TP or only TN reference sets.

  • 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.

    My main concern is that the paper’s “state-of-the-art” framing is too strong and not sufficiently contextualized. The experiments show that RGB outperforms the selected loss-function baselines under the authors’ own protocol, but this is not the same as establishing broader state of the art on CXR-LT. The CXR-LT 2024 challenge paper reports that Team A took first place in Task 1 with mAP 0.281, mAUROC 0.847, and mF1 0.289, and it specifically notes that Team A also achieved the best overall performance and strong tail-group performance. This matters because the RGB paper’s reported CXR-LT result, 79.33 mAUC, is numerically below the official challenge-leading CXR-LT 2024 result of about 0.847 mAUROC, even before considering stronger system design choices. While the metric naming and exact protocol are not perfectly identical, the burden is still on the authors to delimit the scope of their claim. As written, the paper gives the impression of broader SOTA than the evidence supports.

    A related issue is that the paper does not position itself against stronger CXR-LT system-level methods such as CheXFusion, which was the 1st-place solution in the ICCV CVAMD 2023 Shared Task on CXR-LT and reported 0.372 mAP on the MIMIC-CXR test set using multi-view transformer fusion, balancing, and self-training. CheXFusion and the 2024 challenge-winning pipeline are not directly comparable line-by-line to RGB because they use different task years, label sets, and system architectures. However, that is exactly the point: RGB is being evaluated as a loss-only method in a fixed DenseNet121 setup, not as a top-performing CXR-LT system. Therefore, the paper should not claim broad state of the art on CXR-LT without a much more careful qualification.

    More broadly, the validation scope is somewhat narrow. All experiments are limited to chest X-ray datasets and a single DenseNet121 backbone. There is no evidence that the method transfers to other medical modalities, stronger architectures, or multi-view/study-level settings, which are especially relevant in the CXR-LT literature. The method is also batch-dependent, since the reference-guided correction relies on the presence of TP or TN reference predictions within a batch. The paper provides a fallback when TP is absent, but it does not study sensitivity to batch size or extremely sparse class prevalence, which seems important for the tail classes it aims to help.

    Finally, the mechanistic support for the “dominant negative gradients” claim remains somewhat limited. The gradient visualization is suggestive, but the paper would be stronger with explicit statistics on gradient magnitudes for head, medium, and tail classes over training, instead of mainly qualitative figures.

  • 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?

    This is a reasonable and practically motivated loss-design paper, and the proposed RGB loss appears to improve over the selected prior losses under the authors’ own setup. However, the broader performance framing is overstated. In light of the official CXR-LT 2024 challenge results, where Team A achieved first place with roughly 0.281 mAP / 0.847 mAUROC / 0.289 mF1, and prior challenge-winning systems such as CheXFusion, the current submission does not establish overall state of the art for CXR-LT. I would recommend that the authors substantially soften the SOTA claim and position RGB as a strong loss-function improvement rather than a leading end-to-end CXR-LT system.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The rebuttal partially addresses my concern by acknowledging that the “state-of-the-art” claim should be narrowed to a loss-function comparison rather than broader CXR-LT system-level performance. However, this also confirms my main concern from the original review: the current paper framing overstates the evidence. The submitted experiments show that RGB improves over selected loss-function baselines under a fixed DenseNet121 protocol, but they do not establish overall state of the art on CXR-LT or against stronger system-level CXR-LT methods.

    Several important concerns also remain unresolved. The rebuttal states that gamma sensitivity, batch-size behavior, and ResNet50 results were omitted due to space. These analyses are directly relevant because the proposed method is batch-dependent and relies on TP/TN reference predictions within a batch. While the authors’ explanations are reasonable, they are not sufficient to resolve the concern without evidence in the submitted manuscript.

    Overall, I view RGB as a useful and practically motivated loss-function improvement, and the reported gains over the selected baselines are promising. However, the novelty and performance claims remain too strong relative to the validation scope. Therefore, I maintain my negative recommendation unless the paper is substantially reframed as a loss-level contribution rather than a broadly state-of-the-art CXR-LT system.



Review #3

  • Please describe the contribution of the paper

    The paper proposes Reference-Guided Gradient Balancing (RGB) loss for long-tailed multi-label medical image classification. RGB comprises a positive enhancement loss that weights gradients by sample difficulty and class frequency, and an adaptive negative suppression loss that leverages correct predictions (true positives or true negatives) within a batch as references to rectify false positive logits. The method is evaluated on three chest X-ray datasets (CXR-14, CXR-LT, CheXpert) and achieves consistent macro F1 and mAUC improvements over nine baseline loss functions.

  • 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.

    S1.Well-motivated reference mechanism. Using correct predictions (TP/TN) within a batch as class-specific references to guide false positive logit correction (Eq. 4) is a novel and intuitive idea that goes beyond uniform negative suppression in DR [14]. The conditional fallback to TN when no TP exists is a practical design choice validated by Table 3. S2.Consistent SoTA across three large-scale datasets. RGB achieves the highest macro F1 on all three datasets (33.27%, 22.45%, 48.02%) and the highest mAUC on CXR-14 and CheXpert (Table 1). Notably, tail-class macro F1 is the best across all datasets, directly supporting the paper’s claims about long-tailed effectiveness.

    S3.Thorough ablation and analysis. The ablation study (Table 2) clearly isolates the contributions of L_pe and L_ans, and the reference set comparison (Table 3) justifies the TP-or-TN design. The gradient visualization (Fig. 2a) provides direct evidence that RGB selectively reduces false positive gradients rather than uniformly suppressing all negatives.

  • 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.

    W1.Reliance on noisy within-batch predictions. The reference loss L_ref (Eq. 4) partitions samples into TP/FP/TN sets based on the model’s own predictions (ŷ), which are unreliable during early training and for tail classes with few positive samples. The paper does not discuss how ŷ is thresholded, how the method behaves when predictions are mostly incorrect, or whether a warm-up period is needed. This is a fundamental design concern that deserves analysis (e.g., tracking reference set accuracy over epochs).

    W2.All three datasets are chest X-rays. While the datasets differ in scale and label count (14/37/13 classes), they share the same imaging modality. Generalizability to other MIA modalities (CT, fundus, dermoscopy, histopathology) that have different positive/negative ratios and label structures remains undemonstrated. The paper’s title and claims are broader than the experimental evidence supports.

    W3.Edge case in λ (Eq. 2) and missing hyperparameter sensitivity. When D_k = 1, log_2(D_k) = 0 and λ is undefined (division by zero); when D_k = 2, λ = 0, which nullifies L_pe for that class entirely. This is problematic for exactly the rare tail classes the method targets. Additionally, while γ1 and γ2 require per-dataset grid search (3/3 for CXR-14, 2.4/3 for CXR-LT, 1.9/2 for CheXpert), no sensitivity analysis is provided beyond the claim of stability within ±0.1. W4.Overclaimed novelty. Contribution (1) states RGB is “the first multi-label loss specifically designed to address the inherent dominance of negative gradients in MIA,” but ASL [19] explicitly applies asymmetric focusing to suppress negative gradients, and RAL [18] directly addresses positive-negative gradient asymmetry in long-tailed multi-label settings. The claim should be scoped to the reference-guided mechanism specifically.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    1.The mi-F1 column in Table 2 shows that L_c-lsep + L_ans drops mi-F1 substantially on CXR-14 (38.58 → 26.69) and CXR-LT (51.59 → 28.00), indicating L_ans alone can harm instance-level performance. Please discuss this trade-off between macro and micro metrics.

    2.Table 1 shows RGB’s macro recall on CXR-LT (24.01) is lower than RAL (26.34) and LSEP (8.79 but with much higher precision). Reporting per-class F1 for at least the tail classes would help readers understand where RGB succeeds and fails.

    3.Please clarify how the prediction ŷ_{i,k} is obtained in Eq. 4 — is a fixed threshold of 0.5 used on the sigmoid output, or is the threshold adaptive? This choice directly affects the quality of the TP/FP/TN partition.

  • 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 reference-guided mechanism is a novel and well-motivated approach to negative gradient dominance in long-tailed multi-label classification, with consistent SoTA results on three datasets. However, the reliance on noisy batch predictions without stability analysis, the λ edge case for extreme tail classes, limited modality diversity, and overclaimed novelty collectively prevent a stronger recommendation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

We sincerely appreciate the reviewers’ valuable comments. The major concerns and our responses are listed below. The source code will be released upon acceptance. λ and γ (R1,R3): We agree that λ is more precisely a class-specific weighting coefficient rather than a dynamic parameter, and we will revise the wording accordingly. While D_k denotes the number of images per class, we verified commonly used multi-label medical image datasets prior to proposing this method, and to the best of our knowledge, D_k ≤ 2 does not occur, including for tail classes. While γ sensitivity analysis was omitted due to space, performance remains robust across a wide range of values. SOTA / Novelty claim (R2,R3): We agree that RGB focuses on loss function design rather than system-level optimization. Our intention is to demonstrate its effectiveness relative to representative multi-label loss functions, not claim overall CXR-LT challenge leadership. Accordingly, we will revise our SOTA statement to a multi-label loss design in the revision. Regarding novelty, we agree that prior methods, such as ASL and RAL, also suppress negative gradients and will remove Contribution (1) to better focus the other three contributions, as the reviewer suggests. Batch-dependent (R2,R3): We agree that RGB is batch-dependent. However, the conditional “TP or TN” design mitigates data-sparse issues by using TN references when TP is absent. Empirically, RGB shows stable gains as batch size increases from 64 to 256 (512 was memory-limited). This analysis was omitted due to space limitations. Strong performance on the highly sparse CXR-LT dataset further suggests robustness in extreme long-tailed settings. Prediction definition (R1,R3): We will clarify the probability p in the revision. The threshold of prediction ŷ_{i,k} is fixed to 0.5 for all experiments, and we will clarify this point in the revision. Problem definition (R1): We value the reviewer’s observation on the category scale. While CXR-LT contains 37 labels, it was specifically designed for long-tailed problems in MIA with a heavy-tailed distribution. This formulation is standard in recent MIA literature, including the CXR-LT MICCAI challenge. Generalizability (R2,R3): We agree that evaluation on different modalities is a valuable future direction. As a loss-level formulation operating solely on logits, RGB is directly compatible with different architectures and MIA modalities without modification. We also evaluated RGB on ResNet50 (to be released with the code) and observed a consistent performance trend. Due to space limitations, the results were not included in the submission. Gradient Conflict (R1): While L_sup suppresses TN logits, L_ref further uses them as relative anchors to guide FP logits to the correct side of the decision boundary via pairwise logit comparison. Thus, TN logits serve as a reference structure rather than being explicitly optimized upward. This synergy is empirically validated by our “TP or TN” strategy, which achieves the best performance across datasets (Tab. 3). Partition (R1): We determine cut points based on the “elbow” point in the descending prevalence distribution. We will clarify this setting in the revision. No correct predictions & warm up (R3): In practice, TN references are inherently abundant because negative samples dominate in sparse medical datasets, making cases with no correct predictions rare. Confidence weighting naturally filters early-training noise, removing the need for a warm-up stage. Trade-off between metrics (R3): L_ans improves ma-F1 by mitigating dominant-negative gradients, while slightly shifting optimization from head-class-dominated mi-F1 in L_clsep. The proposed L_pe solves this and shows our overall design enhances both macro and micro metrics (Table 2), confirming that the two components work complementarily. Figures (R1): We will export the vector image format in the 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.

    Reviewers are split: one recommends weak reject and two recommend weak accept. After careful consideration, I recommend sending this paper for rebuttal. The reference-guided mechanism is novel and well-motivated — using TP/TN predictions as class-specific anchors to guide false positive suppression is an elegant idea, and consistent improvements across three chest X-ray benchmarks are encouraging. The ablation studies are thorough and the method is easy to integrate as a pure loss replacement. However, several concerns need to be addressed. The SOTA claim is overstated — RGB is evaluated as a loss-only method under a fixed DenseNet121 setup and does not establish broader state of the art relative to challenge-winning systems on CXR-LT. The method relies on within-batch predictions as references, which are unreliable during early training and for extreme tail classes, yet no stability analysis is provided. A mathematical edge case exists in the λ formulation that breaks for the rarest classes (D_k ≤ 2), which is precisely the setting the method targets. Evaluation is limited to a single modality and backbone, limiting generalizability claims.

    The authors are asked to clarify the following points in their rebuttal:

    • How the λ edge case is handled for classes with D_k ≤ 2.- The stability of within-batch reference predictions over training, particularly for tail classes.
    • Whether the SOTA claim can be scoped more precisely to loss-function comparisons rather than system-level results.
    • How ŷ is thresholded in the TP/FP/TN partition.
  • 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.

    This paper introduces a well-motivated algorithm that addresses a related problem. However, a key concern is the clarity of its claimed contribution, which the authors have agreed to revise during the rebuttal phase. Overall, I believe the paper is worthy of discussion at MICCAI.



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.

    This paper received Accept, Reject, and Weak Accept (no post-rebuttal submitted).

    The authors’ rebuttal agreed to narrow the SOTA claim to a loss-function comparison and to remove Contribution (1), but the framing still overstates the evidence and the validation of the batch-dependent design is not in the submitted paper.

    All three reviewers agreed the loss formulation is practically motivated and the gains over selected baselines are real, and the Reject is conditional on the framing reframe the rebuttal commits to.

    The authors should narrow the SOTA claim, remove Contribution (1), and discuss stronger system-level CXR-LT baselines in related work.



Meta-review #3

  • 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.

    After rebuttal, the reviewer opinions remain mixed. Reviewer #1 supports acceptance, stating that most concerns were addressed, while Reviewer #2 maintains rejection because the SOTA claim remains over-scoped and important evidence on batch-size behavior, γ sensitivity, and stronger backbones is absent from the submitted manuscript. Reviewer #3 does not provide a visible post-rebuttal final opinion in the provided text, but the initial review was weak accept with concerns about batch-dependent references, λ edge cases, and limited modality/backbone diversity.

    Overall, the proposed RGB loss is a practical and reasonably novel loss-level contribution with consistent gains over selected loss-function baselines, but it should not be framed as a broadly state-of-the-art CXR-LT system. The final version clearly narrows the claims to loss-function comparison and explicitly discusses the remaining limitations regarding batch dependence, tail-class stability, λ edge cases, single-backbone evaluation, and limited modality generalization. The authors should provide the full open-source implementation to improve the reproducibility.



back to top