Abstract

Data scarcity is a major bottleneck in medical Multiple Instance Learning (MIL), especially for rare diseases and expensive modalities. We introduce a statistically grounded patient augmentation approach that generates realistic patients directly in embedding space. Using Gaussian mixture models to probabilistically cluster pooled instance embeddings from all patients, our method learns disease-specific “recipes”—statistical distributions of instances across unsupervised clusters. New patients are then generated by sampling embeddings from clusters based on learned recipes. Unlike existing methods that require examples from all categories, our method can generate patients offline by remixing pooled embeddings. Generated patients are further selected based on uncertainty quantification to improve MIL performance. We evaluate our method across three clinically relevant scenarios: (i) cross-dataset transfer, where an entirely missing healthy class is generated using statistics from an external cohort; (ii) low-data regimes, where class sizes are extremely limited; and (iii) small-cohort non-image tasks, including single-cell RNA-seq and flow cytometry. Across all experiments, our method improves performance over baseline, often outperforming other bag-mixing strategies. Notably, in the missing-class scenario, a performance comparable to full-dataset training is achieved, demonstrating its potential for rare disease diagnostic and privacy-preserving patient augmentation. The code is available at https://github.com/marrlab/RECIPE.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3998_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/marrlab/RECIPE

Link to the Dataset(s)

AML_Hehr: https://www.cancerimagingarchive.net/collection/aml-cytomorphology_mll_helmholtz/ cAItomorph: https://nefeli.helmholtz-munich.de/records/9bv4e-3ag16 Covid-flow: http://flowrepository.org/id/FR-FCM-Z5PC

BibTex

@InProceedings{DasMuh_Remixing_MICCAI2026,
        author = { Dasdelen, Muhammed Furkan AND Ozlugedik, Fatih AND Litinetskaya, Anastasia AND Navab, Nassir AND Marr, Carsten AND Sadafi, Ario},
        title = { { Remixing Embeddings for Patient Augmentation in Data Scarce Multiple Instance Learning } },
        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

    This paper addresses the data scarcity issue in medical multi-instance learning (MIL) and proposes a novel and statistically motivated patient-level data augmentation method. The core idea is to use the Gaussian Mixture Model (GMM) to cluster the instances in the embedding space, learn “disease-specific recipes”, and generate synthetic patients through re-mixing the embeddings. This method demonstrates superior performance over the baseline in three clinically relevant scarcity scenarios: cross-dataset transfer, low data regimes, and non-image tasks (such as single-cell RNA-seq, flow cytometry).

  • 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 experimental scenarios are diverse and challenging, including three rare scenarios cover the absence of categories, a very small number of labeled samples, and non-image modalities, which verify the generalization ability of the method. 2.The proposed method in the article is novel. Instead of performing simple interpolation or replication at the original data level, it conducts “re-mixing” in the embedding space, which better conforms to the bag structure characteristics of MIL. By using GMM to learn the clustering distribution and then performing resampling, it offers more statistical interpretability compared to existing bag mixture methods.

  • 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.GMM clustering is performed on pooled instance embeddings. When the number of patients or the number of instances in each bag is large (such as patches in WSI), what are the complexities of GMM training and sampling? Please provide data such as training time and GPU memory consumption, and discuss whether the method is applicable to extremely large MIL datasets (such as millions of instances). This is very important for clinical deployment.

  • 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.

    (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?

    Please refer to the major strengths.

  • 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.

    N/A



Review #2

  • Please describe the contribution of the paper
    • Addresses the challenge of Multi-Instance Learning (MIL) where only bag-level labels are available without instance-level annotations.
    • Proposes a novel method to synthesize artificial patients (bags) by recombining existing instances.
    • Utilizes Gaussian Mixture Models (GMMs) on instance embeddings to cluster similar instances.
    • Derives bag-label “recipes,” defined as the stochastic distribution of different clusters for each label in the dataset.
    • Applies uncertainty estimation to the generated bags, selecting only high-uncertainty (hard) samples for downstream training.
    • Demonstrates the ability to generate bags for a disease class absent from the primary dataset using a recipe learned from an external dataset.
    • Shows that these synthetic bags improve downstream classifier performance, particularly for zero-shot or underrepresented 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.
    • Successfully generates instance bags for out-of-distribution classes based on external datasets, producing distributions similar to real data.
    • Achieves impressive results (near 1.0) for previously unavailable or reduced classes.
    • Evaluates the proposed method across three distinct datasets.
    • Employs multiple relevant baselines for comparison.
    • Presents a wide variety of results with extensive overall analysis.
  • 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.

    Statistics and Validation Strategy:

    • Lack of a Held-Out Test Set: The methodology appears to rely solely on a validation split, which severely limits the validity of the reported results. Hyperparameters (e.g., K=50, augmentation percentage, uncertainty method, inclusion criteria) were optimized during this process. Given that the performance differences compared to baselines are small relative to the standard deviations, it is impossible to determine whether the gains stem from the proposed method or mere hyperparameter overfitting. A held-out test set or nested cross-validation is necessary.
    • Missing Statistical Tests: No statistical significance tests are provided. While means and standard deviations are reported, the results appear to be marginally better only in isolated cases. The standard deviations frequently overlap, especially in the AML experiments (which constitute the majority of the evaluation).

    Missing Baselines and Ablations:

    • External Data Baseline: Generating recipes for an unseen class from an external dataset assumes the availability of that external data. The authors must include a baseline where this external data is simply included directly in the downstream training to isolate the true benefit of the recipe-generation approach.
    • Component Ablation: The pipeline consists of two distinct modules: sample generation and sample selection (uncertainty filtering). The individual contribution of each module is not established. An ablation demonstrating performance when using all generated samples (without uncertainty exclusion) is missing.
    • Inconsistent Ablation Setup: The ablation study is performed on a different dataset than the main AML experiments. Furthermore, two different types of embeddings are used for the same dataset during ablation, which is counter-intuitive.
    • Cluster Size (K) Selection: During the ablation, K=3 appears to be the optimal number of clusters. However, K=50 (which is absent from the ablation table) is chosen for the main experiments because it allegedly “provided the best resolution for capturing semantic diversity.” The authors fail to explain or quantify how this semantic diversity was assessed.

    Evaluation Shortcomings and Generalizability:

    • Class Imbalance in AML: The AML experiments only evaluate the generation of the healthy class. To support claims of generalizability, experiments generating other classes are necessary. The confusion matrix shows that NPM1 and RUNX1 are underperforming; an experiment upsampling these specific classes is missing.
    • Implausible Baseline: The baseline confusion matrix in Fig. 3 seems implausible, showing near-random performance across all classes even though only the healthy cases were excluded from training. One would expect the model to fail primarily on the excluded healthy cases, not universally.
    • Misleading Selection Claims: The claim that “Global selection achieved the highest bAcc (0.73 ± 0.09)…” is misleading, as all approaches (except “most certain”) achieve very similar results well within the margin of error (high standard deviation).
    • An upper Performance Limit is missing, where samples are generated till Saturation (additional Figure: generated samples vs metric)

    Missing Implementation Details:

    • It is not stated how the recipes are applied to the non-AML datasets (e.g., how many samples are created, and are the classes balanced after generation?).
    • The rules and thresholds for the uncertainty-based sample inclusion/exclusion are not defined. The exact number of samples retained versus discarded is also missing.
  • Please rate the clarity and organization of this paper

    Poor

  • 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 provide sufficient information for 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
    • Caption Overload: Image and table captions are overly extensive. Major methodological details should be moved from the captions into the main text to improve readability.
    • Figure 2 Clarifications:
    • Explain how UMAPs are calculated at the bag level.
    • Add a clear sub-caption to (D) stating that the left is generated and the right is the baseline.
    • The table in (E) should be formatted as a proper table in the text, or removed if the information is redundant with Figure 3.* Figure 3: The confusion matrices require numerical values for proper interpretation.
    • Focus: Due to page limitations, the authors should consider focusing the detailed methodology and ablations entirely on one dataset (AML) and briefly summarize the results for the others.
    • Rebuttal Advice: Given time constraints, the authors should prioritize defending and clarifying their most unique contribution: the generation of completely new bags from external recipes.
  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    While the authors address a highly relevant problem (data scarcity in MIL) and introduce an new method to derive label compositions from external datasets, the manuscript lacks necessary evaluation. The absence of a dedicated test split and the optimization of parameters on the validation set make the small performance gains statistically questionable. Furthermore, essential baselines—such as directly training on the external data used for the recipe—are missing. The analysis lacks the necessary ablations to justify architectural choices, and fundamental implementation details regarding sample generation and uncertainty filtering are omitted.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.
    • While the authors addressed some minor remarks, they failed to resolve major weaknesses, primarily the lack of statistical significance. Given the high standard deviations reported in the main results and ablation studies, the actual efficacy of the proposed method remains unclear.
    • The rebuttal clarifies the use of a held-out test set, a critical detail omitted from the manuscript. However, the cross-validation methodology and the calculation of the standard deviation remain ambiguous. It is unclear whether the variance is computed across the validation sets or on a fixed held-out test set using different train/val splits.
    • The ablation study is restricted to the COVID-19 dataset, yet the main results emphasize the AML dataset. Furthermore, the ablation results show no significant variance across most configurations. For example, Aug % margins are all within +/- 1% with a reported standard deviation of ~0.15.Because Aug % is intended to be a core parameter, this lack of variance fails to isolate the contributions of individual modules or prove the overall effect of the method.
    • The authors state that COVID-19 hyperparameter results were omitted due to space constraints. This hampers reproducibility and reflects odd prioritization of manuscript space. Notably, the selection of the second-highest parameter value ($K=50$) is highly questionable, as Table 2 indicates a clear trend where smaller $K$ values yield better results. This discrepancy is not justified.
    • The unexpected and catastrophic failure of the baseline results in Figure 3 remains entirely unexplained.
    • The organization of the methods and results is convoluted. Different components are evaluated on disparate datasets without a cohesive structural logic.
    • Neither the manuscript’s discussion nor the rebuttal addresses the method’s lack of significant performance improvements or its overall insensitivity to different parameter configurations.



Review #3

  • Please describe the contribution of the paper

    The paper introduces a framework how to re-mix embeddings in a multiple instance learning classification setup for digital pathology. The proposed approach comprises the training of a Gaussian mixture model, sampling from existing patches to generate new patients and choosing the most informative patients for training based on entropy as uncertainty measure. The approach is model and data agnostic and shown to outperform related mixup strategies in their specific experiments.

  • 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) The proposed methodological framework (i.e. employment of Gaussian mixture model, sampling from existing data points to craft artificial patients, Monte Carlo dropout to gain uncertainty predictions) appears sound and well-justified.

    S2) The experimental validation is comprehensive and convincing. Different datasets from both, imaging and flow cytometry data domains, are used; and related mixup variants are employed as a comparison to existing literature ablation studies are included.

  • 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 find no major methodological flaws; the main issues are related to clarity and presentation.

    W1) the experiments (e.g. in table 1) state “mixup”, “remix” and “psemix” as other mixup-variants to be tested. However, only remix and psemix are cited in the introduction. What particular mixup variant was used for vanilla mixup?

    W2) The dataset descriptions in section 3.1.name include a dataset named PBMC, which is described as a “single-cell RNA-seq cohort”. Experiments in Table 1 and 2 refer to the dataset as “scRNA-seq” or “Covid-19 scVI/PCA embeddings; it would be helpful to use more consistent naming.

    W3) In section 3.2., last paragraph (“For each real patient […]”): Since n_{c,k} is drawn from a continuous distribution but is meant to be the (discrete) number of instances used for a re-mixed patient, the number should also be discretized somehow. (e.g. use floor or ceiling symbols)

  • 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.

    (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?

    Overall, this is a solid and well‑executed contribution, with a sound methodological foundation and comprehensive experimental validation. The identified weaknesses are minor and primarily concern clarity and presentation, and can be readily addressed.

  • 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.

    My initial review only contained a few minor comments, all of which have been addressed in the rebuttal.

    I maintain my original recommendation.



Author Feedback

We thank the reviewers for their feedback and recognition of our contributions, including the novelty of the study (R1-3), sound methodology (R3) and extensive evaluation scenarios (R1-3). Below, we clarify the points discussed by the reviewers.

R1: Complexity of GMM fit – We used scikit-learn (Section 3.2), which optimizes the expectation-maximization objective for data log-likelihood on CPU. Complexity is O(n) with diagonal covariances (Section 3.2), and convergence depends on quality, dimensionality, and number of embeddings. As an example, fitting on 650K embeddings in scenario 1 takes 10 minutes on a single CPU. Notably, the model needs fitting only once. We make sure to clarify this advantage in the paper.

R3: Consistent naming, notations and missing MixUp reference – We apologize for inconsistent dataset names, along with the missing ceiling operation in recipe formation per patient which are now fixed. MixUp in our experiments was the original interpolation-based version, implemented at slide level by Liu et al., 2024.For this reason, we cited PseMix (Liu et al., 2024) and will also cite the original MixUp paper.

R2: Lack of held-out test set and use of different embeddings for ablations – We believe there is a misunderstanding: i) We used 5-fold cross validation with a held-out test set in all experiments (Section 3.1). ii) Hyperparameter selection used a single dataset with one embedding type (Table 2, Covid-19 scRNA-seq scVI); the same hyperparameters were used for all other datasets/experiments.

R2: Missing external data baseline – For scenario 1 (where an entire class was missing and the missing-class recipe was derived from an external dataset), we already included a baseline in which this external data was directly used in the downstream task, as suggested by the reviewer. We showed a clear benefit of patient generation over direct transfer (see Figure 2E). We will improve the clarity and highlight this point.

R2: Component ablation (uncertainty selection is missing) – We would like to clarify that this point is already addressed in Table 2.We select patients based on their uncertainty to help the model distinguish challenging cases. We already benchmarked different selection strategies among generated patients, including the least uncertain patients, random patients, and the most uncertain patients. Adding the most uncertain patients to the dataset improved model performance the most.

R2: Cluster size K selection – We would like to clarify a misunderstanding. Given limited space, the ablation was condensed into Table 2.Hyperparameters used in the study are given as default setting in the caption. In each section of the hyperparameter column in Table 2, we varied one parameter while keeping the rest the same as the default. For cluster size, we ablated K=3,10,30,50 (default), 100.The best configuration used K=50, not K=3. R2: Missing upper performance limit, where samples are generated till saturation – We would like to note that the limit is not missing, but shown in Table 2 where we ablate the added ratio of generated patients. Baseline (0% generated patients) achieves 0.58 balanced accuracy (Table 1), including generated patients consistently improves: 0.68, 0.69, 0.70, 0.73 for adding 5, 10, 20 and 30% generated patients respectively (Table 2). Using more patients did not improve after this point.

R2: Missing how many samples are created, rules and thresholds for the uncertainty – Thanks for highlighting this point. For scenarios 1 and 2, patient counts are given in Section 4, page 5.For scenario 3, we generated the same number as the original dataset and selected the most uncertain 30% (see Table 2). No threshold was used; generated patients were sorted based on their uncertainty and most uncertain patients were selected based on the specified ratio. We will clarify this for scenario 3. The source code will be released in a public GitHub repository.




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 received mixed reviews, including two accept and one reject. The authors are invited to submit a rebuttal to address the reviewers’ concerns. I recommend inviting a rebuttal because the paper receives clearly mixed reviews and has potentially strong methodological value. Two reviewers recommend acceptance, highlighting that the proposed GMM-based patient-level re-mixing strategy is novel, statistically motivated, and well suited to the MIL setting. They also recognize the diversity of evaluation scenarios, including cross-dataset transfer, low-data regimes, and non-image modalities. However, one reviewer raises substantial concerns about validation strategy, missing held-out testing, lack of statistical tests, incomplete baselines, insufficient ablations, unclear uncertainty filtering, and missing implementation details. A rebuttal would allow the authors to clarify the evaluation protocol, justify the choice of K, explain sample generation and selection rules, provide or discuss statistical significance, and address whether external-data recipes offer benefits beyond directly adding external data. Given the disagreement among reviewers and the possibility that several concerns can be clarified or partially addressed, a rebuttal is warranted before the final decision.

  • 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 rebuttal, this paper receives two accept recommendations and one rejection. The AC acknowledges that the paper addresses an important data scarcity problem in medical MIL and proposes a novel GMM-based patient-level re-mixing strategy that is statistically motivated and well aligned with bag-structured data. The reviewers recognized the method’s originality, its ability to synthesize patient-level bags from learned disease-specific recipes, and its evaluation across diverse settings including cross-dataset transfer, low-data regimes, and non-image modalities. The rebuttal clarifies several major concerns, including the existence of held-out test folds within cross-validation, the external-data baseline, the uncertainty-based selection strategy, cluster-size selection, generated-sample ratios, and computational complexity. Although concerns remain regarding statistical significance, high variance, and clarity of the experimental organization, the majority of reviewers consider the methodological idea sound and potentially valuable for data-scarce MIL. Therefore, the AC recommends acceptance.



Meta-review #2

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

    Accept

  • Please justify your recommendation.

    The paper proposes RECIPE, a patient-level augmentation framework for data-scarce medical multiple-instance learning. The method fits a Gaussian mixture model to pooled instance embeddings, derives class-specific cluster-distribution “recipes,” generates synthetic patient bags by re-mixing embeddings according to these recipes, and selects informative generated patients using uncertainty estimates. A particularly interesting aspect of the work is the ability to generate a missing patient class using an externally derived recipe, which is relevant for rare-disease and data-scarce medical settings.

    The initial reviews were mixed, with scores of 5 / 2 / 5.Reviewers 1 and 3 appreciated the novelty and practical relevance of patient-level embedding re-mixing, the ability to generate missing-class bags from external statistics, and the broad evaluation across imaging and non-image MIL settings. Reviewer 2 raised important concerns regarding validation strategy, statistical significance, hyperparameter selection, ablation design, and missing implementation details.

    The rebuttal addressed several of the major concerns. In particular, the authors clarified that all experiments used 5-fold cross-validation with a held-out test set, and that hyperparameters were selected on a single dataset and embedding type, Covid-19 scRNA-seq with scVI embeddings, and then fixed for all other datasets and experiments. The authors also clarified that the direct external-data transfer baseline was already included, that uncertainty-based sample selection was ablated in Table 2, that K=50 was the default and best configuration in the ablation, and that generated samples were selected by ranking uncertainty rather than by applying an additional threshold.

    Some limitations remain. The manuscript should more clearly describe the cross-validation protocol, the train/validation/test split within each fold, and how the reported standard deviations were computed. The ablation study is mainly performed on the Covid-19 scVI setting rather than on the AML missing-class setting that motivates the paper most strongly. Statistical significance testing is not provided, and the reported gains should therefore be interpreted with some caution. The final version should also clarify the hyperparameter-selection protocol, uncertainty-based selection rules, and the scope of the claims regarding generalization to missing or underrepresented disease classes.

    After rebuttal, the reviewer opinions remain mixed but overall positive, with two reviewers recommending Accept and one reviewer maintaining Reject. I place weight on the concerns raised by Reviewer 2, but I find that the rebuttal sufficiently mitigates the most serious issues, particularly the concerns about the absence of a held-out test set and hyperparameter selection. The remaining issues are important limitations and clarity/reproducibility concerns, but they do not invalidate the main contribution.

    I recommend Accept. The paper presents an interesting and useful idea for patient-level augmentation in data-scarce MIL, provides evidence across multiple clinically relevant scenarios, and offers a promising approach for missing-class or low-sample settings. The final manuscript should temper claims, clarify the validation and hyperparameter-selection procedures, and explicitly discuss the remaining limitations regarding statistical significance and generalization.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    The paper addresses the problem of extreme sample scarcity in whole-slide histopathology classification, where traditional patient-level data collection is restricted. To solve this, the authors propose a generative augmentation approach that models the latent feature space via Gaussian mixtures to mix and sample novel synthetic patient representations. The authors demonstrated strong performance boosts in data-scarce and cross-dataset transfer scenarios. The rebuttal resolved most concerns regarding uncertainty filtering mechanics and provided thorough hyperparameter ablations over cluster sizes, and validated the stability of the generative limits. The methodological value supports the acceptance of the paper.



back to top