List of Papers Browse by Subject Areas Author List
Abstract
Semi-supervised domain generalization (SSDG) has recently emerged as an appealing alternative to tackle domain generalization when labeled data is scarce but unlabeled samples across domains are abundant. In this work, we identify an important limitation that hampers the deployment of state-of-the-art methods on more challenging but practical scenarios. In particular, state-of-the-art SSDG severely suffers in the presence of long-tailed class distributions, an arguably common situation in real-world settings. To alleviate this limitation, we propose \textbf{\ours{}}, a simple yet effective objective based on the well-known InfoMax principle adapted to the SSDG scenario, where the Mutual Information (MI) between the learned features and latent labels is maximized, constrained by the supervision from the labeled samples. Our formulation integrates an $\alpha$-entropic objective, which mitigates the class-balance bias encoded in the standard marginal entropy term of the MI, thereby better handling arbitrary class distributions. \ours{} can be seamlessly plugged into recent state-of-the-art SSDG, consistently enhancing their performance, as demonstrated empirically across two different image modalities.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0191_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FilLeo_Information_MICCAI2026,
author = { Fillioux, Leo AND Chakraborty, Omprakash AND Gopée, Quentin AND Marza, Pierre AND Cournède, Paul-Henry AND Christodoulidis, Stergios AND Vakalopoulou, Maria AND Ben Ayed, Ismail AND Dolz, Jose},
title = { { Information Maximization for Long-Tailed Semi-Supervised Domain Generalization } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Authors present Imax, method for semi-supervised domain generalisation that is tuned for imbalanced data scenarios. They derive optimisation objective based on mutual information principles. They compare the method with recent SSDG algorithms, FBCSA and DGWM. They show improved classification for two sets of bio-medical images, when domain shift, class imbalance, and low number of labeled images have to be considered.
- 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.
New method for classification in the SSDG scenario when classes are imbalanced. The method shows competitive performance on this type of data when compared to recent SSDG algorithms. This is achieved by additional entropy terms in the optimisation objective. The method could be of general interest.
- 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 comparison to the relevant state of the art, in particular SSL methods for imbalanced data sets. The manuscript does not show how the method differentiate from the recent SSL methods for imbalanced datasets. There are several such methods (for example Adsh, https://proceedings. mlr. press/v162/guo22e/guo22e. pdf?utm_source=chatgpt. com), and they are not included neither in the introduction nor in the methods comparison. SSDG methods FBCSA and DGWM were tested with FixMatch, StyleMatch, and FreeMatch SSLs (and augmented with Imax), but in principle, Adsh (or related methods) could also be used instead, correct? Please clarify why such methods were not included.
Limited clarity of presentation. Some parts of the manuscripts need additional clarifications, as enumerated below.
SSL methods such as FixMatch already optimise the crossentropy loss and pseudo crossentropy loss. Does Imax duplicate these by including the two last terms in Equation 6? For the presented results obtained with e. g. FixMatch + Imax, was only the first term from Equation 6 used or all of them? Please clarify.
Should we expect the Imax improvements to be similar for higher m values and thus higher accuracies? These would be more practical settings for medical applications where high accuracy is usually required. Please discuss.
Is the imbalance forced on the labeled, unlabeled, or both data types for the presented experiments? It is not compleately clear from the description. If only on labeled, given m=5, do all classes have labeled members?
- 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
A comparison for a more general audience on some standard sets (such as PACS) would be useful, if available (no additional experiments are requested).
Parentheses (+5.0% and +3.4%) seem to be duplicated in the results.
- 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?
Based on (limited) literature search, Imax seems to be the first one or one of the first methods for SSDG in the context of imbalanced data. The problem is clearly presented, the MI-based approach is motivated, and the method comparison shows improvement over the two selected SSDG methods. However, several works on SSLs for imbalanced data are not considered here but are likely relevant, since the main SSDG methods can switch between SSL approaches.
- 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 address my main concern, i.e., the lack of comparison with existing SSL methods for imbalanced data, by including LA and CB losses in comparisons. While these methods serve as a baseline, they do not represent the state of the art in imbalanced SSL. In my opinion, inclusion of some recent methods, such as BEM, CReST (which can be combined with LA), or Adsh would be more informative regarding the presented method performance. Therefore, I am not fully convinced that the presented method is an improvement over the combination of existing approaches for DG and imbalanced SSL.
Review #2
- Please describe the contribution of the paper
This paper proposes iMax, a semi‑supervised domain generalization (SSDG) method based on the Information Maximization (InfoMax) principle. The approach maximizes mutual information between learned representations and latent labels, while being constrained by limited labeled data from source domains. To address long‑tailed class imbalance, the method further integrates an α‑entropic objective, encouraging balanced feature label associations across classes. The method is evaluated on two medical imaging modalities: histopathology and ophthalmology (diabetic retinopathy), Clear and well‑written presentation: The paper is easy to follow, with a clear motivation, formulation, and experimental setup. Novel adaptation of InfoMax to SSDG: Applying mutual‑information maximization to the semi‑supervised domain generalization setting is technically interesting and well motivated. Relevance to medical imaging: Domain generalization under limited labels is a critical challenge in medical AI, making the proposed approach practically relevant. Handling of class imbalance: The integration of an α‑entropic objective to mitigate long‑tailed distributions is a meaningful addition. Strong experimental comparisons: The paper compares against relevant SSDG and DG baselines and reports consistent improvements across tasks.demonstrating improved generalization performance under limited supervision and domain shift.
- 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) Clear and well‑written presentation: The paper is easy to follow, with a clear motivation, formulation, and experimental setup. 2) Novel adaptation of InfoMax to SSDG: Applying mutual‑information maximization to the semi‑supervised domain generalization setting is technically interesting and well motivated. 3) Relevance to medical imaging: Domain generalization under limited labels is a critical challenge in medical AI, making the proposed approach practically relevant. 4) Handling of class imbalance: The integration of an α‑entropic objective to mitigate long‑tailed distributions is a meaningful addition. 5) Strong experimental comparisons: The paper compares against relevant SSDG and DG baselines and reports consistent improvements across tasks.
- 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) Inconsistent behavior across modalities: In Table 1, for the retinal task, simpler baselines combined with iMax outperform FBCSA+iMax and DGWM+iMax, while the opposite trend is observed for the histopathology (ESCA) dataset. The paper does not provide sufficient explanation for this modality‑dependent behavior.
2) Insufficient dataset details: The paper does not clearly report the number of images per dataset, nor how many are used as labeled, unlabeled, and test samples for each modality. It is also unclear whether all target domains contain comparable numbers of samples.
3) Limited reporting of variability: Table 1 reports only the mean performance across four target domains. Standard deviations are missing, and it would be helpful to report the best‑ and worst‑case target domains, along with the number of samples in those domains.
4) Limited analysis at higher label budgets: Performance gains appear to diminish already at m_L = 10.It would strengthen the paper to evaluate higher labeled budgets (e.g., m_L=20, 50) to assess whether iMax continues to provide benefits when baselines are already strong.
5) Clarity missing in Figure 1: The meaning of the gray and green bars on the left side of Figure 1 is unclear and not sufficiently explained in either the caption or the main text.
- 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 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?
The paper presents a technically sound and well‑motivated approach to semi‑supervised domain generalization, with a novel InfoMax‑based objective and thoughtful handling of class imbalance. The experimental results across two medical imaging modalities are promising, and comparisons to prior work are generally strong. However, the paper would benefit from clearer dataset descriptions, more detailed statistical reporting, and deeper analysis of observed performance differences across modalities and label budgets. These issues primarily affect clarity and interpretation rather than the core validity of the method and can likely be addressed with relatively minor revisions.
- 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 clarifications on class imbalance handling, evaluation metrics (including class-wise performance), and the behavior across datasets are helpful and address my main concerns. In particular, the explanation that modality-dependent trends pre-exist in the baseline methods (and are consistently improved by iMax) strengthens confidence in the robustness of the proposed approach. The additional justification of the α‑entropic objective and its role in stabilizing the mutual information objective under long-tailed SSDG settings is also seems reasonable. Overall, the concerns raised in the initial review were primarily related to clarity and interpretation rather than methodological soundness, and the rebuttal sufficiently addresses these points. I support acceptance.
Review #3
- Please describe the contribution of the paper
The paper introduces IMaX, a plug-and-play information maximization objective for long-tailed semi-supervised domain generalization. The method adapts mutual information maximization to the semi-supervised setting by combining supervised constraints from labeled samples with pseudo-label supervision from unlabeled samples. It further replaces the standard marginal entropy term with an α-entropy objective to reduce the class-balance bias and better handle imbalanced class distributions. The proposed objective is integrated into existing SSDG frameworks and evaluated on histology and retinal image classification tasks.
- 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 a realistic and important limitation of current semi-supervised domain generalization methods, namely their sensitivity to long-tailed class distributions. This setting is particularly relevant for medical imaging, where labeled data are often scarce and clinically important categories may be underrepresented.
A strength of the work is that the proposed objective is simple and easy to integrate into existing SSDG pipelines. Rather than requiring a new network architecture, IMaX can be plugged into different semi-supervised learning frameworks such as FixMatch, FreeMatch, and StyleMatch, as well as recent SSDG strategies including FBCSA and DGWM.
The motivation behind the method is also clear. The paper identifies the class-balance bias induced by the standard marginal entropy term in mutual information maximization, and proposes an α-entropy based alternative to make the objective more tolerant to non-uniform class distributions.
The experimental evaluation covers two medical imaging modalities, histology and retinal imaging, and reports consistent improvements across several combinations of SSL and SSDG methods. This supports the claim that the proposed objective is not tied to a single baseline and has some degree of model-agnostic applicability.
- 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 evidence for the long-tailed setting is not fully convincing. The paper mainly reports overall accuracy, which may be dominated by head classes and does not show whether the proposed method actually improves minority or clinically important classes. For this problem setting, balanced accuracy, macro-F1, per-class sensitivity/recall, head/tail class breakdowns, and confusion matrices would make the evaluation much stronger.
The comparison with imbalance-aware baselines is insufficient. Although methods such as class-balanced loss, Balanced Softmax, logit adjustment, or distribution-alignment based pseudo-labeling are not specifically designed for SSDG, they are standard approaches for handling long-tailed or imbalanced classification. Including them under the same SSL/SSDG framework would help clarify whether the gains come from the proposed information-maximization objective rather than from simpler class-prior correction strategies.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper studies an important and realistic problem, and the proposed IMaX objective is simple, well motivated, and consistently improves several SSL/SSDG baselines. However, the technical novelty is moderate, and the current evaluation does not sufficiently support the long-tailed claim. In particular, the paper mainly reports overall accuracy, without enough minority-class analysis or comparisons with standard imbalance-aware baselines. These limitations make the contribution promising but not yet fully convincing for acceptance.
- 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
Review #4
- Please describe the contribution of the paper
Semi-supervised domain generalization methods typically assume a uniform distribution of data categories in the source domain; however, this assumption is often unrealistic in real-world scenarios. Therefore, a plug-and-play objective function, termed IMaX, is proposed based on the InfoMax principle. The proposed method maximizes the mutual information between learned features and latent labels.
- 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.Building on the InfoMax principle in information theory, the method integrates cross-entropy and pseudo-label consistency from semi-supervised learning into the conditional entropy term of mutual information. 2.The proposed IMaX functions as an auxiliary loss module that can be readily integrated into mainstream semi-supervised learning frameworks. 3.Experimental results demonstrate significant performance improvements.
- 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 treats α as a fixed hyperparameter selected via the validation set. In the SSDG setting, there is a lack of in-depth discussion on how α can be dynamically set or estimated when the source-domain imbalance is unknown or varies over time. 2.Although the task is formulated as SSDG, IMaX primarily addresses the long-tail problem. It remains unclear why the introduction of Tsallis entropy improves performance on unseen target domains. Consequently, the paper lacks sufficient analysis in this aspect.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is more suitable for a workshop venue. However, it lacks an in-depth analysis of the theoretical and empirical relationship between the hyperparameter α and the imbalance coefficient γ. Providing further clarification on this aspect in the rebuttal would strengthen the overall evaluation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank reviewers and the AC for the constructive feedback. Due to the limited space, we address below the main points retained by the AC. 1.Clarification on evaluation: the imbalance in our setting impacts only the labeled samples, while the unlabeled samples are not subsampled, reflecting the original class distribution, which is a common setting [1,2]. Consequently, the overall accuracy is not dominated by head classes. We further report complementary metrics to prove superior performance when training with IMaX.
- Class-wise lowest accuracy: this shows the performance among the most difficult classes, key in the context of DG: IMaX yields +6.0 percentage points (pp) higher than models trained without on average (across settings), with higher values across all settings.
- Accuracy for classes with the fewest test samples: IMaX yields +5.4 pp on average, again with higher values among all settings. 2.Differences from standard imbalance-aware baselines: logit adjustment (LA) and class-balance (CB) losses are designed for standard long-tailed tasks. In a general setting, they work well as we have access to the marginal class distributions. However, this is incompatible with the long-tailed SSDG setting for two reasons: labeled-set class frequencies are known, but cover only a minority of the data, and the distribution in the labeled and unlabeled set are very different, which will introduce a bias towards the data. Neither addresses the implicit assumption about the uniformity of the data in the mutual information objective, which is relaxed with IMaX. We implemented both losses into our framework: they score on average 3.5 pp and 30.33 pp lower than IMaX, respectively, with IMaX winning in all settings. 3.Tab. 1: Baselines + IMax outperform FBCSA/DGWM+IMaX in the retinal dataset, while the opposite holds for histology data (ESCA). We stress that this trend is already present without IMaX, i.e., IMaX does not introduce this behavior. In contrast, IMaX consistently improves the underlying method (in all but one setting), with substantial margins. The observed trend likely stems from dataset specificities: FBCSA/DGWM rely on domain-specific modulation, which helps when domain shift dominates (ESCA), but is less helpful under more severe class imbalance and weaker domain variability (retina). 4.How to fix α? Fig. 2 addresses this issue by showing that validation and test accuracies follow similar trends. The value of α can therefore be chosen on the validation set, which is always accessible, as is standard practice in machine learning. Our experiments suggest that values of α∈[1.5,2] lead to robust gains. 5.Theoretical explanation of α-entropy improvement for domain generalization. We would like to recall that the MICCAI submission guidelines state that “the evaluation and performance assessment of presented works can potentially be limited to proofs of concept or small-scale validation studies”. Thus, we believe that we should not be penalized for not providing theoretical results. Furthermore, the statement (R4) “although the task is formulated as SSDG, IMaX primarily addresses the long‑tail problem” does not accurately reflect our formulation. IMaX is derived as a semi-supervised mutual-information objective for SSDG (Eqs. 2-8), and is always applied on top of SSDG methods under the standard SSDG protocol. The α-entropy is not designed to replace DG mechanisms, but to stabilize the MI objective under long-tail distributions, thereby improving the robustness of SSDG methods to domain shift. Indeed, Tsallis entropy improves generalization to unseen domains because it relaxes the overly rigid uniform-marginal constraint of KL-based MI, yielding more stable class-marginal estimates across domains, which is essential for SSDG. [1] SimPro: A simple probabilistic framework towards realistic LT semi-supervised learning. ICML’24 [2] Towards realistic long-tailed semi-supervised learning: Consistency is all you need. CVPR’24
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.
Two reviewers recommend weak accept and two recommend weak rejection. After careful consideration, I lean toward weak rejection and recommend sending this paper for rebuttal. The paper addresses a practically important problem — semi-supervised domain generalization under long-tailed class distributions — and the InfoMax-based objective is well-motivated, simple to integrate, and shows consistent improvements across multiple SSL/SSDG frameworks. However, several concerns need to be addressed. The evaluation relies primarily on overall accuracy, which may be dominated by head classes and does not convincingly demonstrate improvement on minority classes. The most significant issue is the lack of comparisons with standard imbalance-aware baselines, which makes it unclear whether the improvements are truly due to the proposed objective. The modality-dependent behavior in Table 1 is unexplained. The authors are asked to clarify the following points in their rebuttal:
- Whether existing results in the paper already support minority-class improvements, and if per-class breakdowns can be derived from reported data.
- How the method conceptually differs from standard imbalance-aware baselines such as logit adjustment or class-balanced loss.
- The theoretical or intuitive explanation for the modality-dependent behavior in Table 1.* How α should be set when source-domain imbalance is unknown, based on existing analysis.
- Why Tsallis entropy is expected to improve generalization to unseen target domains from a theoretical standpoint.
- 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 main concerns regarding this paper relate to the need for clarifications, its connection to imbalanced learning baselines and imbalanced semi-supervised learning (SSL) baselines, as well as the rationale behind the design choices. In the rebuttal, the authors provided clarifications and partially addressed these concerns. Overall, the strengths of the paper outweigh its weaknesses, and I therefore recommend acceptance.
To further strengthen their contributions, I suggest that the authors include additional imbalanced SSL baselines, such as Adsh and CReST. I also recommend adding more discussion on the use of Tasllis entropy (α-entropy) compared to other entropy formulations. This would help position INFOMAX more effectively within the imbalanced SSL literature (see, for example, Mey, Alexander, and Marco Loog. “Improved Generalization in Semi-Supervised Learning: A Survey of Theoretical Results.” IEEE Transactions on Pattern Analysis and Machine Intelligence 45.4 (2022): 4747–4767).
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 introduces a technically sound and model-agnostic objective that successfully tackles long-tailed class imbalances under semi-supervised domain generalization. Through a convincing rebuttal, the authors effectively addressed initial critiques regarding the lack of minority-class analysis and standard baseline comparisons. Considering the robust empirical gains demonstrated across both histology and retinal imaging tasks, I recommend the acceptance of this manuscript.
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.
I recommend reject: although the rebuttal addressed several concerns by AC, adding LA/CB comparisons and per-class metrics, R3 and R4 did not engage with the rebuttal and their Weak Rejects (3) stand on the moderate technical novelty. The central claim that IMaX outperforms imbalance-aware alternatives is not adequately substantiated against current SOTA.