acceptodds
Withdrawn from ICLR 2027

Label-Free Coreset Selection with Foundation Models for Efficient Annotationin Computational Pathology

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Computational pathology has the potential to improve clinical outcomes through a demonstrated increase in diagnostic and prognostic accuracy. However, the development and validation of deep learning algorithms still require annotated data, a costly procedure involving expert pathologists who already face critical workforce shortages. Existing coreset selection methods to optimize annotation efforts currently all rely on hyperparameters tuned on natural-image benchmarks that do not transfer to histopathology and are cumbersome to use in clinical practice. In this study, we present GCcore, a novel label-free coreset selection method that embeds every image of a dataset with any pathology foundation model and greedily selects the samples that collectively maximize the global coverage of the embedding space. The proposed method provides a lower-bound guarantee on the global coverage of the returned coreset for any coreset size, while being completely hyperparameter-free and deterministic. We demonstrate GCcore's superior performance over 14 baselines including state-of-the-art methods across 10 tasks and datasets spanning whole slide image classification, tile classification, and tissue segmentation, where it ranks first on six and within the top three on nine, while also demonstrating how existing methods can shift by up to five rank positions depending on their hyperparameter settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Coreset Selection, Histopathology, Deep Learning, Foundation Model, Label-Free Coreset Selection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cNHSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GCcore, which uses pathology foundation-model embeddings to select representative subsets of unlabeled data for efficient annotation and downstream learning through deterministic, hyperparameter-free greedy maximization of global coverage.

Critical strengths

1.The paper addresses an important annotation-efficiency problem in computational pathology, with broad evaluation across datasets, organs, and task families, including segmentation. The study focuses on practical annotation budgets of tens to hundreds of samples.

2.For fixed embeddings and distances, GCcore offers a simple, deterministic procedure without clustering or weighting hyperparameters, with consistently competitive empirical performance. The results also suggest robustness to encoder choice and class imbalance.

Critical weaknesses

1.GCcore’s objective is equivalent to classical facility location, while its approximation guarantee and acceleration follow established results. The claimed methodological novelty is therefore overstated, and the manuscript does not clearly acknowledge this equivalence.

2.Averaging baseline performance over an author-chosen configuration grid makes the reported rankings dependent on that grid and does not sufficiently establish practical superiority. These averages need not reflect performance under reasonable default settings.

Questions for authors

What new methodological contribution does this work provide beyond the established selection framework?

2 comments
AuthorsOpenReview ↗

Authors' Responses

Dear Reviewer cNHS,

We sincerely appreciate your recognition of our work and your constructive comments. We hope the revisions address your concerns, and we would be grateful if you would consider them in your final assessment. If any concerns remain, please let us know, and we will be glad to address them.

Weakness-1: Equivalence to classical facility location and methodological novelty.

Thank you for pointing this out. We agree with this comment and acknowledge that the original manuscript did not state the equivalence clearly enough. We have revised the manuscript to state it explicitly (Section 3, new paragraph “Relation to facility location”). We have also narrowed our claims to what is new throughout the Abstract and Sections 1, 3, and 5. The revised content is highlighted in green.

Weakness-2: Author-chosen hyperparameter configurations of baselines may not reflect performance under reasonable default settings.

Thank you for this comment. To ensure a fair comparison, we did not choose the baseline configurations arbitrarily. For each baseline, the five hyperparameter values are sampled within the range reported to be robust in the original paper, where such a range is given, and otherwise around the reported default value (Section 4.1, paragraph “Baselines”). In both cases, the default value is included. For example, InfoMax (Tan et al., 2025) was reported to be robust to its pairwise weight over α∈[0.3,4]\alpha \in [0.3, 4], so we used α∈0.3,1,2,3,4\alpha \in \\{0.3, 1, 2, 3, 4\\}. Averaging over these values therefore reflects the performance a practitioner can expect when following the original authors' recommendations without labeled data for tuning, which is the setting of label-free coreset selection. A detailed explanation is provided in Appendix F.1, ”Hyperparameter settings of baselines.”

Questions for Authors: What new methodological contribution does this work provide beyond the established selection framework?

Thank you for the question. As stated in the new paragraph “Relation to facility location” in Section 3, our methodological contribution is to show that this classical objective arises from, and is equivalent to, a radius-free measure of dataset coverage, providing a new, intuitive interpretation of the objective. Starting from the fraction p(r;T)p(r;\mathcal{T}) of the dataset lying within radius rr of the coreset, we integrate over all radii to remove the dependence on an arbitrary, scale-specific rr. Using the tail-integral identity (Eqs. 5-7), we prove that the resulting measure, GC(T)\mathrm{GC}(\mathcal{T}), equals the normalized facility location function F(T)/(Nrmax⁡)F(\mathcal{T})/(N r_{\max}). To our knowledge, this interpretation has not been made explicit before. It has two practical consequences. First, it justifies the objective in directly interpretable terms: maximizing FF is equivalent to maximizing coverage averaged over every radius scale, rather than at a single hand-tuned radius. Second, it provides a normalized score GC∈[0,1]\mathrm{GC}\in[0,1] that can be compared across FMs, datasets, and budgets. Combined with the label-free, FM-based pipeline, this makes GCcore, to our knowledge, the first application of this coverage-maximizing framework to annotation-efficient computational pathology.

Thank you for reading.

Thank you for the authors' response. I decide to keep my original score.

Reviewer_7s75OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GCcore, a label-free coreset selection method for efficient data selection in computational pathology. It uses a pretrained pathology foundation model to embed images and greedily selects samples to maximize global coverage of the embedding space, defined as the area under the coverage–radius curve. Given the coreset size, the method is deterministic and requires no hyperparameter tuning. The authors show that the objective is monotone and submodular, yielding a ((1-1/e)) approximation guarantee for greedy selection. They evaluate GCcore against 14 label-free baselines on 10 datasets covering whole-slide image (WSI) classification, tile classification, and tissue segmentation. The downstream models use frozen foundation-model encoders with trainable task-specific heads. In the low-retention regime, GCcore ranks first on six datasets and among the top three on nine. The paper also investigates the sensitivity of competing methods to hyperparameter configurations.

Critical strengths

GCcore selects samples by greedily maximizing the global coverage of the embedding space. It is hyperparameter-free and deterministic, given the coreset size. The authors also provide a lower-bound guarantee on global coverage. These properties make the method practical when expert annotations are costly and labels are unavailable for hyperparameter tuning.

Critical weaknesses
  1. Equation 7 establishes that maximizing global coverage is equivalent to minimizing the mean nearest-coreset distance. The approximation guarantee is based on a standard result for monotone submodular maximization. However, the paper does not sufficiently clarify the methodological novelty of the proposed objective and optimization method beyond their application to computational pathology. A clearer explanation of the methodological contributions relative to existing approaches is needed to assess the novelty and significance of the work.
  2. The main comparison reports the average performance of each baseline across five hyperparameter configurations. While this evaluation is relevant in settings where labels are unavailable for hyperparameter tuning, the resulting rankings may depend on the selected configurations and their ranges. Moreover, the reported standard deviations (SDs) primarily reflect sensitivity to hyperparameter configurations rather than variability across independent training runs. Reporting results for individual configurations, together with variability across repeated runs, would help establish whether the reported performance gains are robust and consistent.
Questions for authors

1.Equation 7 expresses global coverage in terms of the mean nearest-coreset distance. What is novel about the proposed objective and its greedy optimization compared with prior work? Please clarify which aspects of the method and theoretical analysis constitute the main contributions beyond existing approaches. 2.Could you report the performance under each hyperparameter configuration and the variability across the five downstream training runs described in Appendix F.2? Do the main conclusions remain robust when default hyperparameter settings are used and variability across training runs is taken into account?

2 comments
AuthorsOpenReview ↗

Authors' Responses (Part 1)

Dear Reviewer 7s75,

We greatly appreciate your insightful comments. We have considered each of your suggestions and revised the manuscript accordingly. We hope that these revisions address your concerns and would be grateful if you would take the improvements into consideration in your final assessment. Please let us know if you have any further concerns, and we would be pleased to address them.

Weakness-1: Clarify the methodological novelty of the proposed method beyond their application to computational pathology

Thank you for pointing it out. We agree that maximizing GC is equivalent to minimizing the mean nearest-coreset distance, that this objective is an instance of the classical facility location function, and that the (1−1/e)(1-1/e) guarantee follows from the standard result for monotone submodular maximization. To acknowledge this and clarify our methodological contribution, we have added a new paragraph, “Relation to facility location,” in Section 3. The revised content is highlighted in green.

Weakness-2: Report variability across independent runs in addition to hyperparameter configurations.

Thank you for the suggestion. To ensure a fair comparison between GCcore and the baselines, each hyperparameter configuration of a baseline selects its own coreset, and the task-specific head is trained five times on the selected coreset. Each baseline is therefore evaluated over 5 configurations × 5 independent training runs. The same five-run protocol is applied to the coreset selected by GCcore (Appendix F.2). We agree that reporting the variability across independent runs separately from the variability across configurations would further strengthen the credibility of our method.

AuthorsOpenReview ↗

Authors' Responses (Part 2)

Questions for Authors-1: What is novel about the proposed objective and its greedy optimization compared with prior work?

Thank you for this question. We have clarified in the revised manuscript which aspects of the method and analysis are established and which constitute our contributions.

Regarding the objective and its optimization, we do not claim novelty. As stated in the new paragraph “Relation to facility location” in Section 3, GC is a normalized facility location function, F(T)=Nrmax⁡GC(T)F(\mathcal{T})=N r_{\max}\mathrm{GC}(\mathcal{T}), and the (1−1/e)(1-1/e) guarantee (Nemhauser et al., 1978) and the lazy greedy acceleration (Minoux, 1978) are established results that GCcore inherits.

Regarding the theoretical analysis, our contribution is to show that this classical objective is equivalent to a radius-free measure of dataset coverage, the area under the coverage-radius curve. To our knowledge, this interpretation has not been made explicit before. It has two consequences. First, it justifies the objective in terms that are directly interpretable for coreset quality: maximizing facility location is equivalent to maximizing the fraction of the dataset covered by the coreset, averaged uniformly over every radius scale rather than evaluated at a single hand-tuned radius. Second, it provides a normalized score GC∈[0,1]\mathrm{GC}\in[0,1] that can be compared across FMs, datasets, and budgets.

Questions for Authors-2: Report the performance across five runs for each hyperparameter setting. Do the main conclusions remain robust when accounting for variability across runs under the default setting?

Thank you for the great question. In our experiments, the five hyperparameter values of each baseline are sampled within the range reported to be robust on natural-image benchmarks when such a range is available, and otherwise around the reported default value (Section 4.1, paragraph “Baselines”). In both cases, the default value is included. In the latter case, we sample values around the default rather than using the default alone, because we found that the default value can yield among the worst test performance. For example, on PANDA at a 0.5% retention rate, the default value M=6M = 6 (the number of classes in PANDA) for Proto-Easy results in a test performance 70% lower than that obtained with M=20M = 20 (Table 12). As our analysis accounts for variability across both hyperparameter configurations and independent runs, the main conclusions remain robust.

In the revised manuscript, we have added Appendix G.7 (“Sensitivity of baselines to hyperparameter configurations exceeds variability across independent training runs”). Table 12 reports the test QWK on PANDA at a 0.5% retention rate for every configuration of every baseline, including all five independent runs. Comparing the two sources of variability, we find that hyperparameter configurations influence test performance considerably more than independent training runs do. For all fourteen baselines, the SD across configurations exceeds the average SD across runs by a factor of 3.13 to 14.70 (median 5.64). Because labels are unavailable for tuning before annotation, this sensitivity translates into a risk of poor coreset selection. GCcore has no hyperparameters to tune, so it avoids this risk.

Thank you for reading.

Reviewer_cuymOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GCcore, a label-free coreset selection method for reducing annotation costs in computational pathology. It uses pretrained pathology foundation models to extract image embeddings and greedily selects representative samples by maximizing global coverage. The authors provide a submodular optimization guarantee and evaluate the method against 14 baselines across 10 histopathology datasets covering classification and segmentation tasks.

Critical strengths
  1. Reducing expert annotation costs is a meaningful challenge in computational pathology. The proposed label-free selection framework has potential practical value, particularly when annotation budgets are limited.

  2. The paper evaluates multiple pathology datasets, foundation models, and downstream tasks. The analysis of hyperparameter sensitivity provides useful insights into the reliability of existing selection methods.

Critical weaknesses
  1. Previous studies have extensively explored embedding-based tissue clustering, contrastive representation learning, and label-efficient annotation in histopathology [1–3]. Although GCcore uses modern foundation models and a different selection objective, the paper does not sufficiently explain its conceptual advances over these earlier approaches. The main contribution appears to be an application of established representation-based selection principles rather than a substantially new methodology.

[1] "Deep Contrastive Learning Based Tissue Clustering for Annotation-free Histopathology Image Analysis." Computerized Medical Imaging and Graphics, vol. 97, article 102053, 2022. [2] "Towards Better Understanding and Better Generalization of Few-Shot Classification in Histology Images with Contrastive Learning." International Conference on Learning Representations (ICLR), 2021. [3] "Human-Machine Interactive Tissue Prototype Learning for Label-Efficient Histopathology Image Segmentation." International Conference on Information Processing in Medical Imaging (IPMI), pp. 679–691, 2023.

  1. Most baselines focus on generic coreset selection or embedding-based clustering methods. The evaluation does not directly compare GCcore with existing pathology-specific approaches to reducing annotation effort, such as [3]. Furthermore, the proposed global coverage objective is closely related to classical facility location optimization, and its approximation guarantee follows established submodular optimization theory. These issues make it difficult to assess the independent methodological and practical value of GCcore.
Questions for authors
  1. What is the main methodological contribution of GCcore beyond applying established embedding-based selection and coverage optimization to modern pathology foundation models?

  2. How does GCcore differ from the annotation-free clustering and label-efficient learning approaches in [1–3]? Can the authors clarify why these pathology-specific methods are not included in the experimental discussion or comparisons?

  3. Does the proposed method offer advantages beyond the stronger representations provided by modern foundation models? What evidence from the existing experiments best demonstrates a practical improvement over earlier pathology-specific annotation strategies?

3 comments
AuthorsOpenReview ↗

Authors' Responses (Part 1)

Dear reviewer cuym,

Thank you for the detailed and valuable comments. We have carefully considered your suggestions and revised the manuscript accordingly. We hope these changes address your concerns and appreciate your consideration in the final assessment. We would be happy to address any remaining concerns.

Weakness-1: Previous studies have explored label-efficient methods in histopathology. The paper does not sufficiently explain its conceptual advances over these earlier approaches. The main contribution is unclear.

Thank you for pointing us to these works. We agree that the original submission did not position GCcore clearly enough relative to prior label-efficient methods in computational pathology. We have therefore added the following statement to Section 2: “Label-efficient methods have been proposed in computational pathology (Belharbi et al., 2021; Yan et al., 2022; Yang et al., 2022; Pan et al., 2023), but to our knowledge, none of them addresses the problem of label-free coreset selection: given a large unlabeled cohort and a budget of kk annotations, which samples should pathologists annotate?” We have also revised the contributions in Section 1, including a new first contribution that positions GCcore as, to our knowledge, the first systematic study of task-agnostic, label-free coreset selection in computational pathology. The revised content is highlighted in green.

Weakness-2: The evaluation does not directly compare GCcore with existing pathology-specific approaches to reducing annotation effort. The proposed global coverage objective is closely related to classical facility location optimization, and its approximation guarantee follows established submodular optimization theory.

Thank you for pointing this out. We explain why we do not compare GCcore with these pathology-specific methods in our response to Question 2 below. We agree that the proposed objective is equivalent to the classical facility location objective. We have revised the manuscript to state this equivalence explicitly (Section 3, new paragraph “Relation to facility location”). We have also narrowed our claims to what is new throughout the Abstract and Sections 1, 3, and 5.

AuthorsOpenReview ↗

Authors' Responses (Part 2)

Questions for Authors-1: What is the main methodological contribution of GCcore beyond applying established embedding-based selection and coverage optimization to modern pathology foundation models?

Thank you for the question. We clarify our methodological contribution in a new paragraph, “Relation to facility location,” in Section 3. We start from an intuitive notion of representativeness: the fraction p(r;T)p(r;\mathcal{T}) of the dataset lying within radius rr of the coreset. To remove the dependence on an arbitrary, scale-specific choice of rr, we integrate over all radii, obtaining the area under the coverage–radius curve, GC(T)\mathrm{GC}(\mathcal{T}). Using the tail-integral identity, we then prove that this radius-free coverage measure is exactly the normalized facility location objective, F(T)/(N rmax⁡)F(\mathcal{T})/(N\,r_{\max}). To our knowledge, this interpretation has not been made explicit before, and it has two practical consequences. First, it justifies the objective in terms directly tied to coreset quality: maximizing facility location is equivalent to maximizing coverage averaged uniformly over all radius scales, rather than at a single hand-tuned radius. Second, it yields a normalized, scale-invariant score GC∈[0,1]\mathrm{GC}\in[0,1] that can be reported and compared across foundation models, datasets, and budgets.

Questions for Authors-2: How does GCcore differ from the annotation-free clustering and label-efficient learning approaches in [1–3]? And why they are not included in the experimental discussion or comparisons?

Thank you for this question.

Differences from [1–3]. These methods and GCcore address different questions.

[1] eliminates annotation effort by clustering tiles into tissue types. Its output is a tile-level tissue label, so it supports only tasks that can be expressed through tissue composition and cannot produce segmentation masks.

[2] assumes that a small labeled support set is already available and studies how to maximize classifier performance from these labels. It does not address which samples should be labeled in the first place.

[3] asks pathologists to label clusters of tiles rather than individual samples, and trains a segmentation network on the resulting pseudo-masks.

None of these methods addresses the question GCcore is designed for: given a large unlabeled cohort and a budget of kk annotations, which samples should pathologists annotate?

Why they are not included as baselines. Our evaluation protocol isolates the selection step: for every method, the same FM, task-specific head, and training procedure are used, and only the selected subset of kk samples changes. This allows performance differences to be attributed to the selection itself. None of [1–3] fits this protocol.

[1] uses no annotation budget, so it cannot be placed on the retention-rate axis on which all methods are compared. It is also inapplicable to segmentation and to tasks defined by morphology rather than tissue composition, such as IDC versus ILC classification on TCGA-BRCA.

[2] is a learning method rather than a selection method. It specifies how to train on a given labeled set, not which samples to label.

[3] relies on a different form of expert input (cluster-level tissue labels rather than sample-level annotations) and trains on pseudo-labels, so its annotation cost is not measured in the same units as a coreset budget. It is also restricted to segmentation.

AuthorsOpenReview ↗

Authors' Responses (Part 3)

Questions for Authors-3: Does the proposed method offer advantages beyond the stronger representations provided by modern foundation models? What evidence from the existing experiments best demonstrates a practical improvement over earlier pathology-specific annotation strategies?

Thank you for the great question.

Advantages beyond FM representations. In all our experiments, every selection method operates on the same FM embeddings, and the downstream model (frozen FM and task-specific head) is trained identically. The only component that differs between methods is therefore the selection itself, so the performance differences in Tables 1 and 2 are attributable to the selection strategy rather than to the strength of the representation. Under this controlled setting, GCcore achieves the best mean rank across the 10 datasets.

GCcore also offers practical advantages that do not depend on representation quality. It is deterministic and hyperparameter-free, whereas the rank of existing methods shifts by up to 4.5 positions depending solely on their own hyperparameters (Table 1).

Evidence of practical improvement over existing pathology-specific annotation strategies. As discussed in our response to Question 2, the earlier pathology-specific methods [1–3] address different questions and cannot be compared under a common annotation budget. To our knowledge, no prior method in computational pathology addresses which samples to select for annotation from an unlabeled cohort. The clearest practical illustration is the class-imbalanced setting (Appendix G.4): on IMB2, GCcore recovers 90.7% of the full-data performance with only 19 WSIs, a level that random selection reaches only with 751 WSIs, i.e., a roughly 40-fold reduction in the number of slides requiring annotation.

Thank you for reading.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.