List of Papers Browse by Subject Areas Author List
Abstract
Training medical image classifiers on entire datasets is wasteful when labeling or compute budgets are constrained: not all samples contribute equally to downstream accuracy. Active learning reduces annotation cost through iterative querying, but assumes repeated access to an oracle and requires multiple rounds of model training. One-shot geometry-based methods such as facility location avoid retraining but operate on pairwise distances that ignore the local structure of the data manifold.
We propose a graph-based one-shot selection method that prunes a labeled training pool using frozen foundation model embeddings. Given embeddings from a pretrained encoder, we construct a k-nearest neighbor graph over all training samples and derive a two-term coverage kernel from the heat diffusion kernel, capturing both direct and two-hop neighborhood relationships. Greedy facility location on this kernel selects class-balanced subsets that maximize coverage of the data manifold. The two-term kernel matches the full spectral heat kernel in selection behavior while reducing computation to sparse matrix operations with a single hyperparameter.
We evaluate on five MedMNIST datasets spanning histopathology, radiology, and microscopy, comparing against both training-dynamics and geometry-based baselines. Our method achieves the highest balanced accuracy on nine of ten dataset-ratio conditions, with the largest gains on class-imbalanced datasets where global graph construction captures cross-class structure that per-class methods miss, all without any model training during selection. Code is available at https://github.com/zahiriddin-rustamov/graph-coverage-selection.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3438_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zahiriddin-rustamov/graph-coverage-selection
Link to the Dataset(s)
MedMNIST v2: https://medmnist.com/
BibTex
@InProceedings{RusZah_OneShot_MICCAI2026,
author = { Rustamov, Zahiriddin AND Badawi, Nadia AND Damseh, Rafat AND Zaki, Nazar},
title = { { One-Shot Data Selection for Medical Image Classification via Graph Coverage } },
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 proposes a one-shot data selection method for medical image classification based on graph coverage over frozen foundation-model embeddings. The method constructs a global k-nearest neighbor graph, defines a simple two-term coverage kernel using 1-hop and 2-hop neighborhood structure, and applies greedy facility location with per-class budget constraints to select representative subsets. Experiments on five MedMNIST datasets show that the method is generally competitive and often outperforms the compared training-dynamics and geometry-based baselines, especially under low-budget settings and class imbalance.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses a practically relevant problem: selecting informative medical training subsets without iterative active learning or repeated oracle access. 2.The proposed graph-based coverage kernel is simple, intuitive, and computationally efficient, while still showing strong empirical performance. 3.The analysis of global versus per-class graph construction is interesting and provides a plausible explanation for the gains under class imbalance. 4.The experimental section is reasonably thorough within its scope, including multiple datasets, two budget settings, five random trials, and ablations on kernel design.
- 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 main practical motivation is somewhat mismatched with the selection protocol. The paper is framed around reducing annotation cost under limited labeling budgets, but the proposed selection procedure enforces per-class budgets, which appears to require label access at selection time. This makes the method less aligned with a truly pre-annotation one-shot setting than the introduction suggests.
2.The methodological novelty is moderate rather than substantial. The paper presents a clean and effective graph-based refinement of geometric subset selection, but the overall formulation still remains close to existing coreset / coverage-style selection paradigms [14, A, B] rather than introducing a fundamentally new selection principle.
[A] Active Learning Through a Covering Lens, NeurIPS 2022 [B] Generalized Coverage for More Robust Low-Budget Active Learning, ECCV 2024
3.The comparison set is not broad enough to fully establish the paper’s empirical position. While the method is compared against several standard training-dynamics and geometry-based baselines, the experimental evaluation does not include a wider range of recent coreset/data-selection methods, which weakens the strength of the empirical claims.
4.The negative conclusion about training-dynamics methods may be stronger than what the experiments fully support. EL2N, Forgetting, and EVA are evaluated under a specific training setup (e. g. , ResNet-18 from scratch, no augmentation, long training), and it is unclear whether this protocol is equally suitable for those baselines. As a result, the paper may be overstating the general inferiority of training-dynamics selection at low budgets.
5.The method depends heavily on the quality of the frozen embedding backbone, yet encoder generality is not tested. All experiments rely on UNI embeddings, even across datasets from different modalities, and the paper itself shows that when the embedding quality is weak, the benefit of the proposed selector largely disappears. This raises questions about how robust the method is under alternative or less well-matched encoders.
6.The evaluation scope is somewhat limited for making broad claims about medical image classification. All experiments are conducted on MedMNIST-style benchmarks with 2% and 5% budgets, so it remains unclear how the method behaves on more realistic medical datasets, under larger budgets, or in settings with different annotation regimes.
7.Some of the paper’s interpretive claims could be better supported. For example, the argument that global graph construction captures decision-boundary structure that per-class methods miss is plausible and supported by case studies, but the evidence remains largely indirect, relying on benchmark improvements and visualization rather than more direct analyses of boundary coverage or selection behavior.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 simple, practical, and technically sound graph-based one-shot data selection method that shows consistently strong results on the evaluated benchmarks. I particularly appreciate the clean formulation, the computational efficiency of the two-term coverage kernel, and the analysis showing the benefit of global graph construction under class imbalance. My score is tempered by the fact that the novelty is moderate, the comparison set could be broader, and some claims—especially regarding training-dynamics baselines and broader generality—would benefit from stronger validation. Overall, however, I find the paper useful, well executed, and likely to be of practical interest.
- 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 rebuttal partially resolves my concerns. The authors clarified the scope of the method as label-efficient pruning of a labeled pool rather than fully unlabeled active learning, which addresses the apparent mismatch between per-class budgets and the annotation-cost motivation. Their explanation that labels only enter through the partition-matroid constraint, while the graph kernel and marginal-gain computation remain label-free, is helpful. The authors also provide reasonable justification for the selected baselines, explaining why several gradient-based coreset methods are not directly compatible with the one-shot setting, and they clarify that UNI is held fixed across all selectors, so the reported gaps primarily reflect the selection strategy rather than encoder differences.
Some concerns remain only partially addressed. The method is still best viewed as a graph-based refinement of existing coreset/coverage-style selection rather than a fundamentally new selection paradigm. Broader comparisons to recent coverage-based or foundation-model-based selectors, cross-encoder validation, and larger-scale/high-resolution medical datasets remain future work. Nevertheless, the rebuttal makes the scope of the contribution clearer and alleviates several fairness and positioning concerns. Overall, I remain positive about the paper as a simple, technically sound, and practically useful contribution.
Review #2
- Please describe the contribution of the paper
This paper addresses the problem of efficient data selection under limited annotation budgets for medical image classification. The authors construct a k-nearest neighbor graph over all training samples and derive a two-term coverage kernel from the heat diffusion kernel, capturing both direct and two-hop neighborhood relationships while reducing computation to sparse matrix operations. Then, the class-balanced subsets that maximize coverage of the data manifold is selected.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The method is conceptually simple and theoretically grounded. It approximates the heat diffusion kernel via a two-term truncation over a k-nearest neighbor graph, capturing both local and broader neighborhood structure of the data manifold with a single hyperparameter. 2.The method operates entirely on frozen foundation model embeddings with no model training required, reducing computation to sparse matrix operations and avoiding the cubic cost of full eigendecomposition. 3.The method is evaluated on five MedMNIST datasets spanning three imaging modalities (histopathology, radiology, and microscopy), and the improved classification performance demonstrates its effectiveness.
- 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 lacks necessary preliminaries on the heat diffusion kernel and graph coverage, which are central to the proposed method. 2.The analysis at the end of Section 2.2 would be better placed in the experiment section, as it discusses empirical observations rather than methodological details. 3.In the ablation study of Section 3.2, the authors conclude that “additional hops neither reliably help nor hurt,” but this is not well-supported: the results of 5% PathMNIST and 2% OrganAMNIST actually improve with additional hops, and the analysis only explores up to three hops, which is insufficient to support such a conclusion. It remains unclear why the authors consider these improvements unreliable. Similarly, the claim that “two-term kernel matches the full spectral heat kernel in selection behavior” in abstract also lacks sufficient theoretical or empirical support. 4.The paper contains grammar errors, such as the misuse of “demonstrating” in contribution 3 of the introduction section.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 primary factors are the lack of theoretical grounding for the two-term kernel approximation and the unsupported ablation conclusions in Section 3.2.
- 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 rebuttal addresses all four weaknesses.
Review #3
- Please describe the contribution of the paper
The paper proposes a subset selection method extracts embeddings from a frozen deep model and constructs a k-nearest-neighbor graph using cosine similarity. Further, the author gave a kernel that uses single-hop and two-hop neighbours to capture similar datapoint of the graph. Finally, a Facility Location submodular function is applied on this kernel to select a representative subset of the data.
- 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 explores the idea of subset selection in a one-shot data selection scenario using model encoding and graphs. The experiments of the methodology are performed on five datasets with varying numbers of classes. The paper is well written and easy to understand. The paper contains very good ablation studies to support the claims.
- 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.Weak baseline, no comparison given with current SOTA . Suggestion of a few subset selection methods, GLISTER (https://ojs.aaai.org/index.php/AAAI/article/view/16988), CORESET https://proceedings.mlr.press/v119/mirzasoleiman20a , Learning from less data (doi: 10.1109/WACV.2019.00142) , Large-scale Dataset Pruning with Dynamic Uncertainty ( doi: 10.1109/CVPRW63382.2024.00767.), GRAD-MATCH (https://proceedings.mlr.press/v139/killamsetty21a) can be potentially comparabel baseline. 2.No justification is given for the particular model ( UNI ) is used. 3.Section 2.3, facility location is a submodular function; multiple submodular functions can be used that have the coverage property. Why only particular function is used. refer https://doi.org/10.1007/BFb0006528
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 results are not compared against the current SOTA baseline.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Authors does not provide appropriate justification.
Author Feedback
We thank R1, R2, R3, MR for recognizing the method as “simple” (R1, R2), “theoretically grounded” (R2), and “technically sound” (R1), with “very good ablation studies” (R3).
Relation to Zheng 2019 RA (MR): We will cite RA and reframe as a graph-based variant per MR. Three method-level differences remain. (i) Operator: RA uses cosine max-cover on per-dataset AE/GAN/VAE features over ≤1.6K segmentation patches; ours uses K=A_sym+A_sym^2 (Eq. 2) on 165K classification pools via frozen UNI + kNN graph. (ii) Class balance: RA has none; we impose a partition matroid (Eq. 3) with 1/2-approximation guarantee. (iii) Cross-class coverage: our global graph (Tab. 3, §3.3); RA’s two-stage max-cover runs intra-cluster then on merged S_c. R1’s [A] ProbCover, [B] MaxHerding: class-blind low-budget AL; neither uses multi-hop propagation or matroid balance.
Label use (MR, R1): The kernel (Eq. 2) and marginal-gain step (Eq. 3) operate on embeddings only; labels enter solely via the per-class budget — a partition-matroid constraint that, if removed, preserves submodular guarantees. We scope the contribution to label-efficient pruning of a labeled pool, with unlabeled extension as future work. All 8 selectors share per-class budgets, isolating selector quality.
Baselines and SOTA (MR, R1, R3): Tab. 1’s 7 baselines span the one-shot family: FL, FPS, Herding (geometry); EL2N, Forgetting, EVA (training-dynamics); Random. On R3’s refs: GLISTER, GRAD-MATCH, CRAIG (deep) need iterative gradient feedback, incompatible with one-shot (GLISTER also needs a validation set); Dyn-Unc is in the training-dynamics family we evaluate, where our low-budget finding applies; Kaushal 2019 applies submodular FL on CNN features (Wei 2015), an algorithm family Tab. 1’s FL on UNI covers. ENRICH and Ji 2025 (FM-based, §1) lack graph propagation; FL in Tab. 1 represents this property, our heat-kernel adds multi-hop coverage both lack. Submodular variants (R3.3): graph-cut/log-det target diversity, FL targets coverage; we use FL because low budgets favor prototypes over spread. Broader recent SOTA remains future work.
Encoder isolation (MR, R1, R3): We use UNI as it is a public, large-scale histopathology foundation encoder adopted without task-specific tuning. It is held constant across selectors in Tab. 1, so the cross-method gap isolates the selector, not the encoder. Ours improves over FL by +5.2/+3.9pp and Herding by +1.1/+2.9pp (Blood 5%, Path 2%); 9/10 wins (p=0.02). Blood/Organ-A,S are out-of-domain for histopath-trained UNI yet ours wins, supporting graph coverage over encoder fit. TissueMNIST 2% is embedding-bounded (§3.2). Cross-encoder study is future work.
Training-dynamics (R1#4): scores follow each method’s published recipe; downstream training (§3.1, ResNet-18, 1000 epochs, no aug) is held fixed across selectors. The low-budget gap reflects retaining hardest examples being suboptimal when data is scarce (Sorscher 2022, §3.3), not selector-protocol mismatch.
Kernel and hops (R2): §2.2 motivates K via the heat-kernel polynomial expansion ∑ c_k A_sym^k (Kondor & Lafferty 2002); correlation >0.999 between the two-term and full spectral kernel (Tab. 2) supports two-term truncation. On hops: no depth wins universally: 1-hop on Path 5%, 2-hop on Path 2%/OrganA 5%, 3-hop on OrganA 2%; max gap 1.6pp. We drop the “neither reliably help nor hurt” phrasing accordingly.
Evaluation scope (MR, R1): We use 224×224 MedMNIST+ (not 28×28) across 5 datasets / 3 modalities (radiology, microscopy, pathology). We address long-tail (DermaMNIST 58:1, global-graph +5.7/+3.4pp, Tab. 3), heterogeneity (3 modalities), and scale (TissueMNIST 165K; sparse 89s vs dense FL >102GB, §3.3). High-resolution and cross-institution shift are scoped as future work.
R1#7 (boundary): direct boundary-coverage analysis is future work.
R2 minor: we will add heat-kernel preliminaries (W1), move §2.2’s empirical analysis to Results (W2), and fix “demonstrating” in contribution 3 (W4).
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.
1.The submission appears closely related to a 2019 paper [ref1]. In particular, both works are built around the same central idea: replacing iterative active learning with a one-shot, coverage-oriented selection strategy that operates in a learned feature space to choose representative, non-redundant samples before downstream model training. While the present paper differs in its specific instantiation—using frozen foundation-model embeddings, a k-NN graph, and a graph-based coverage kernel for classification rather than unsupervised feature extraction plus clustering/max-cover for segmentation—the overall methodological paradigm and selection philosophy remain highly similar. As such, I believe the 2019 paper is highly relevant prior work that should be explicitly discussed, and the current manuscript’s novelty would be better framed as a graph-based variant of an existing one-shot representative selection paradigm rather than as a broadly new direction in medical image data selection.
[ref1] Zheng, H., Yang, L., Chen, J., Han, J., Zhang, Y., Liang, P., Zhao, Z., Wang, C., & Chen, D. Z. (2019). Biomedical Image Segmentation via Representative Annotation. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 5901-5908.https://doi.org/10.1609/aaai.v33i01.33015901
2.A major methodological concern is that the proposed selection procedure is not label-free at selection time. Both the paper and the released code indicate that the method uses ground-truth class labels to impose equal per-class budgets during subset construction. This means the method assumes access to class membership information for the full candidate pool before annotation, which is inconsistent with the standard unlabeled-pool selection setting that the paper appears to motivate. Under this protocol, the task is not truly “which samples should be labeled from an unlabeled dataset,” but rather “which class-balanced subset should be retained from an already labeled dataset.” This distinction is important, because access to labels materially changes the difficulty and realism of the problem. The authors should clarify the exact problem setting they address, moderate claims related to unlabeled-data selection, and preferably include experiments in a setting where class labels are not available during selection.
3.The experimental comparison is not sufficiently strong. While the paper includes several classic baselines such as Facility Location, FPS, Herding, EL2N, Forgetting, and EVA, the overall baseline set appears limited and somewhat dated. In particular, the comparisons do not seem to include enough strong recent methods that are competitive for modern subset selection or foundation-model-based data selection. As a result, the current experiments mainly show improvement over a relatively basic set of baselines, rather than providing convincing evidence that the proposed method outperforms stronger or more up-to-date alternatives.
4.The practical significance of the evaluation setting is somewhat limited. The experiments are conducted mainly on MedMNIST benchmarks, which are lightweight and highly standardized classification datasets. While useful for controlled comparisons, such benchmarks only weakly reflect the complexity of real medical annotation settings, which often involve high-resolution images, stronger heterogeneity, long-tail distributions, domain shift, and more substantial annotation burden. As a result, the current experimental setup demonstrates effectiveness mainly in a simplified benchmark regime, and it remains unclear how well the proposed method would translate to more realistic medical data selection scenarios.
5.The conclusions may be confounded by the specific embedding backbone used for selection. The method is evaluated entirely on frozen UNI embeddings, even though the datasets span multiple modalities beyond pathology. Since the proposed selection mechanism operates only on the embedding geometry, its effectiveness may depend heavily on the representational quality and domain suitability of the chosen foundation model. The paper itself notes that when the embeddings are weak, all methods tend to converge. Therefore, without comparisons across multiple pretrained encoders or modality-matched backbones, it is difficult to determine whether the gains come from the proposed graph coverage principle itself or from a favorable interaction with one particular embedding space.
The rebuttal is expected to directly and convincingly address the concerns outlined above. As it currently stands, the paper has several major flaws in problem formulation, methodology, and experimental validation, which substantially weaken the overall contribution.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I thank the authors for their rebuttal and for providing additional details and clarifications. The responses adequately addressed my concerns, and I therefore recommend acceptance of this paper. I encourage the authors to revise the paper according to the reviewers’ comments and suggestions, as well as the points mentioned in the rebuttal.
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 final rating contains two accept and one reject. The rebuttal has clarified the scope of the contribution and explained several fairness concerns. More results were provided as well. The rebuttal has also provided explanations regarding negative reviewer’s concern about lack of evaluation of several coreset selection methods. The response seems reasonable.
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.
There are 2 Accepts and 1 Reject after rebuttal. The paper proposes a simple and efficient graph-based one-shot data selection method, and reviewers generally found the idea technically sound. However, I still have concerns about the overall contribution. The problem setting is not fully aligned with the annotation-cost motivation, since the selection protocol uses class labels during subset construction. The method also appears closer to a graph-based variant of existing representative/coverage selection methods than a clearly new selection paradigm. In addition, the validation is mainly on simplified benchmarks, and the comparison to stronger recent subset-selection methods remains limited. Given these concerns, I lean toward rejection.