List of Papers Browse by Subject Areas Author List
Abstract
Recent foundation models for histopathology have created a rapidly expanding set of feature extractors (FEs), yet selecting an appropriate FE for a new whole-slide image (WSI) dataset typically requires repeatedly training Multiple Instance Learning (MIL) models, leading to substantial computational overhead. We propose an efficient pre-evaluation framework that ranks candidate FEs without MIL training. Existing representation evaluation metrics are largely designed for single-image inputs and fail to capture the bag-of-instances nature and spatial heterogeneity of WSIs. To address this mismatch, we introduce two WSI-aware metrics—SAM-Cluster and SAM-LP—that incorporate histological structure by leveraging the Segment Anything Model (SAM) to infer region-level organization and evaluate representation quality at the bag level. Experiments on eight public WSI datasets show that our metrics exhibit high rank correlation with the downstream performance of fully trained MIL models, enabling reliable FE selection at a fraction of the usual cost. Code is available at https://github.com/researchanonymous359-ops/sam-wsi-fe-selection.git
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4643_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/juhyeon-ai/WSI-FE-Selection.git
Link to the Dataset(s)
Camelyon16: https://camelyon16.grand-challenge.org/
BRACS: https://www.bracs.icar.cnr.it/
UBC-OCEAN: https://www.kaggle.com/competitions/UBC-OCEAN/data
HistAI Skin-b1: https://huggingface.co/histai
TCGA-GLIOMA: https://portal.gdc.cancer.gov/
TCGA-NSCLC: https://portal.gdc.cancer.gov/
TCGA-RCC: https://portal.gdc.cancer.gov/
PANDA: https://panda.grand-challenge.org/
BibTex
@InProceedings{KimJuh_SAMCluster_MICCAI2026,
author = { Kim, Juhyeon AND Ssemakula, Paul AND Song, Chanjae AND Yi, Mun Yong},
title = { { SAM-Cluster and SAM-LP: Structure-Aware Evaluation Metrics for Selecting WSI Feature Extractors Without MIL Training } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16891},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces two structure-aware pre-evaluation metrics, SAM-Cluster and SAM-LP, for ranking WSI feature extractors (FEs) without running full MIL training. The central idea is to use SAM-derived region masks as pseudo-structural priors so that representation quality is assessed at the region/bag level rather than through global pooling or i. i. d. patch assumptions. SAM-Cluster evaluates unsupervised separability as a ratio of intra-region cohesion to relaxed inter-region separability, while SAM-LP mitigates signal dilution in supervised evaluation by reweighting structural prototypes through an EM procedure. The paper validates the approach on 8 public WSI datasets against 10 candidate FEs, with a ground-truth ranking constructed from 9 MIL aggregators × 5 seeds, and reports that the proposed metrics correlate substantially better with downstream MIL performance than several generic transferability baselines, making FE screening considerably more practical.
- 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.FE selection for WSI pipelines is expensive, yet highly consequential for downstream performance — the paper reports a 20+ percentage-point swing on Camelyon16 between ResNet50 (76.59%) and Virchow2 (97.98%). This is a meaningful problem with clear real-world impact that has been largely underexplored compared to aggregator design.
2.The methodological intuition is well-motivated. Using SAM masks as pseudo-structural regions is a sensible way to move beyond global statistics and incorporate tissue organization into FE evaluation. The SAM-Cluster metric is simple and interpretable, and SAM-LP provides a reasonable mechanism to counter signal dilution under bag-level supervision by reweighting structural prototypes.
3.The experimental scope is a major strength. The study spans 8 diverse public datasets, 10 candidate feature extractors, 9 MIL aggregators, and 5 random seeds per configuration, which gives the downstream “ground-truth” ranking a credible basis and makes the Spearman correlations meaningful.
4.The paper evaluates both unsupervised and supervised settings, and supplements the main results with analyses on sample efficiency (N from 5 to 100), top-k screening, and hyperparameter sensitivity (τ and β). These additional experiments substantively strengthen the practical value of the work.
- 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 empirical gains are meaningful but not uniformly dominant across all datasets. For the supervised metric, Linear Probing ties SAM-LP exactly on UBC-OCEAN (both 0.8303), SAM-LP is actually slightly lower than LP on TCGA-GLIOMA (0.4424 vs. 0.4545), and LogME is the strongest metric on PANDA (0.7430 vs. SAM-LP’s 0.5879). The paper would benefit from an explicit discussion of these failure/near-tie cases and from a characterization of when the SAM-based prior is most beneficial (e. g. , in terms of structural heterogeneity, tumor-to-background ratio, or bag size).
2.The contribution of SAM is not fully isolated experimentally. Prior work such as SAM-MIL [6] has already established that SAM-derived structural priors are useful in WSI analysis — the novelty here is repurposing such priors as a pre-evaluation signal for FE selection rather than as an aggregator component. This repositioning is reasonable, but the paper does not compare against simpler spatial or regionalization alternatives (e. g. , coordinate-based grouping, superpixels, Otsu-based tissue components, or random region prototypes of matched size). Such ablations are necessary to show that the gains are specific to SAM rather than to any form of region-level grouping.
3.The “ground-truth” FE ranking is defined as the average downstream performance across 9 MIL aggregators. This is reasonable as a broad benchmark, but in practice many users select an FE for a fixed downstream aggregator. Correlations with per-aggregator rankings — or at least a stratified comparison between attention-based aggregators (e. g. , ABMIL, TransMIL) and simpler pooling aggregators — would substantially strengthen the practical claim.
4.Several implementation details remain under-specified for exact reproducibility: the SAM variant and backbone used, the WSI magnification/resolution at which masks are generated, the patch-to-mask assignment rule, any mask filtering or merging step, class balancing in the sampled 30-WSI subsets, and the EM stopping criterion.
5.The computational-efficiency argument is largely qualitative. The paper states that metric computation reduces evaluation time from “weeks” to “minutes” and scales O(1) in |Φ|, but provides no concrete wall-clock runtime, GPU/memory usage, or hardware-based comparison for SAM preprocessing versus even a limited MIL search. A simple table quantifying these savings would directly support one of the paper’s main selling points.
6.The EM formulation in SAM-LP (Eq. 7) reweights each prototype by its isolated predictive likelihood P(Y_i | c_{i,k}; θ), while the prototype embeddings c_{i,k} themselves remain fixed throughout the EM loop (only the weights α and the classifier θ are updated). This is structurally closer to a static attention model than to the context-dependent weighting learned by attention-based MIL aggregators. It is therefore unclear whether SAM-LP is most predictive of attention-based downstream behavior, mean-pooling behavior, or both — a targeted per-aggregator analysis would clarify what class of downstream models SAM-LP is actually forecasting.
- 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.
(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?
I lean Weak Accept because the paper tackles an important and under-studied problem in WSI learning, proposes a sensible structure-aware solution, and supports it with a broad empirical study. The practical value is tangible: even moderately reliable top-k screening of feature extractors could save substantial computation in real MIL pipelines, and the reported correlations (average ρ = 0.5348 for SAM-Cluster and ρ = 0.6258 for SAM-LP) are non-trivial given the difficulty of the task.
I do not rate the paper higher for several reasons. The computational advantage, which is one of the paper’s headline claims, is argued qualitatively rather than quantitatively. The role of SAM itself is not fully disentangled from simpler regionalization alternatives, leaving open whether the gains stem from SAM specifically or from any region-level grouping. The ground-truth ranking is averaged across 9 aggregators, which may overstate the metrics’ utility for practitioners who have already committed to a particular aggregator, and SAM-LP’s EM formulation (with fixed prototypes and isolated per-prototype likelihoods) may be better aligned with some downstream aggregators than others.
Overall I find the contribution useful and promising, with the final judgment depending meaningfully on the rebuttal.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal substantively addresses four of my six concerns. Most importantly, the per-aggregator stratification (W3, W6) directly confirms my hypothesis that SAM-LP’s EM formulation aligns most closely with mean pooling (ρ = 0.6078) while remaining stable across pooling and attention-based aggregators (0.5688 ± 0.0226), and shows the averaged ground-truth ranking is well-represented at the per-aggregator level (mean ρ = 0.8205). The structural-complexity characterization (W1) gives actionable guidance on when SAM-LP helps over LP, and the FLOPs-level breakdown for SAM versus FE extraction (W5) makes the efficiency claim concrete rather than rhetorical. The main residual gap is the SAM-versus-simpler-regionalization ablation (W2), which is deferred to a journal extension; the arguments against Otsu, grid, and SLIC are defensible, but a matched random-region-prototype control would still have strengthened the methodological isolation of SAM’s contribution. On balance, the rebuttal converts the qualitative claims into quantitative evidence and engages directly with the methodological critiques, so I am now comfortable above the acceptance threshold.
Review #2
- Please describe the contribution of the paper
This paper addresses the problem of feature extractor (FE) selection for WSI-based MIL pipelines, which currently requires exhaustively training hundreds of model configurations to identify the best-performing FE for a given dataset. The authors propose a pre-evaluation framework that ranks candidate FEs without any MIL training, using only pre-extracted embeddings from a small slide subset. The core contribution is two WSI-aware suitability metrics: SAM-Cluster, an unsupervised metric that evaluates intra-region feature cohesion and inter-region separability using SAM-derived segmentation masks as structural pseudo-labels; and SAM-LP, a supervised metric that replaces naive mean-pooling with an EM-based prototype weighting scheme to address signal dilution from non-diagnostic background patches. The framework is evaluated across 8 public datasets covering diverse cancer types, reporting Spearman rank correlations against ground-truth MIL performance derived from 90 fully trained configurations per dataset.
- 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 problem formulation is practically motivated and well-scoped. The observation that no single FE dominates across datasets, combined with the scaling argument (10 FEs × 9 aggregators × 5 seeds = 450 runs per dataset), makes a compelling case for a cheaper pre-selection step. This is a real bottleneck in computational pathology research that has not received much direct attention.
The use of SAM as a source of structural pseudo-labels is a reasonable design choice. Rather than relying on pathology-specific annotations, the authors leverage SAM’s zero-shot segmentation to approximate histological boundaries, and Fig. 2 provides at least qualitative evidence that the resulting masks are meaningful. The decoupling of SAM inference from the candidate FEs means the segmentation cost is paid once per WSI regardless of how many FEs are being evaluated, which is a practical advantage.
The experimental setup is unusually thorough for this type of paper. Grounding the evaluation in 90 fully trained configurations per dataset (10 FEs × 9 aggregators × 5 seeds) across 8 diverse datasets provides a fairly robust reference ranking and makes the reported Spearman correlations more credible than if they were based on a single aggregator or a much smaller experimental grid.
The sample size analysis is a useful practical contribution. Showing that SAM-LP approaches near-peak correlation at N = 30 while conventional LP requires substantially more samples directly supports the claim that the framework can reduce selection cost without severely sacrificing reliability.
The hyperparameter sensitivity results are also helpful. The fact that τ and β can be fixed globally without dataset-specific tuning is important for a method intended to be used before committing to a full downstream training pipeline.
- 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 performance gains of SAM-LP over standard Linear Probing are not uniformly strong across datasets. In Table 1, SAM-LP underperforms LP on TCGA-GLIOMA and PANDA, and only modestly improves over it on several others. The average improvement is partly driven by large gains on a few datasets, especially TCGA-NSCLC, while the advantage is smaller or absent on some others. The paper does not really discuss when SAM-based prototyping helps most and when it does not. That limits interpretability of the metric in practice.
SAM’s suitability as a structural prior for pathology images is asserted more than systematically validated. Fig. 2 shows a single qualitative example on colorectal tissue, which is visually convincing, but it remains unclear how well SAM masks capture meaningful boundaries in more difficult settings such as diffuse tumors, necrosis, artifacts, or highly heterogeneous slides. Some broader analysis of SAM mask quality or failure modes across the included datasets would strengthen this assumption.
The EM optimization in SAM-LP requires slide-level labels, so SAM-LP is still a supervised metric. This is not a flaw by itself, but the paper should be clearer about the practical label requirements and where SAM-LP is applicable versus where only SAM-Cluster can be used. Relatedly, it would help to know how many labeled slides are needed before SAM-LP becomes reliable in practice.
The paper does not report the actual wall-clock time or computational cost of running the proposed framework end-to-end. The claim that evaluation is reduced “from weeks to minutes” is plausible, but it is not backed by concrete timing numbers. Since efficiency is a major motivation for the work, a simple runtime table would make this part of the argument much more convincing.
The Top-k analysis is useful, but Top-5 is a relatively loose criterion given that there are only 10 candidate FEs in total. Reporting per-dataset Top-1 and Top-2 selection accuracy more explicitly would make the practical selection value easier to judge.
- 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 paper tackles a practical and underaddressed problem in computational pathology with a reasonable technical approach. The experimental foundation is solid: the reference ranking is derived from an unusually large number of trained configurations, and the evaluation spans 8 diverse datasets. The proposed metrics improve over the main baselines in aggregate, and the sample-size analysis supports the claim that the method can be useful in low-cost FE pre-selection. My main reservations are that the benefits of SAM-LP are not equally strong on all datasets, the efficiency claim is not backed by concrete timing numbers, and the Top-k analysis is not as strong as it could be given the relatively small candidate pool. These are meaningful but addressable concerns rather than fundamental flaws, so overall I lean weak accept.
- 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 authors have responded substantively to all major concerns. The computational cost analysis with concrete FLOPs figures clarifies the efficiency advantage over exhaustive MIL training. The explanation of when SAM-LP outperforms LP — specifically on datasets with high structural complexity such as TCGA-NSCLC and BRACS — provides useful and actionable guidance that was missing from the original submission. The aggregator bias concern is addressed by showing that the averaged GT ranking correlates strongly with per-aggregator rankings (mean ρ = 0.8205), confirming the robustness of the ground truth. The justification for using SAM over simpler alternatives is also reasonable. I am raising my score to 5.
Review #3
- Please describe the contribution of the paper
The paper proposes a MIL-free framework for selecting WSI feature extractors on a target dataset. Its key contribution is two structure-aware metrics, SAM-Cluster and SAM-LP, which use SAM-derived regions to evaluate feature quality at the bag level and better reflect WSI spatial heterogeneity than standard global metrics.
- 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 important problem: feature extractor choice clearly affects downstream MIL performance and currently requires expensive trial-and-error training. (2) The proposed metrics are well motivated for WSI data, especially by incorporating region-level structure instead of relying on naive global pooling or patch-wise i.i.d. assumptions. (3) The empirical study is relatively broad, with multiple datasets, feature extractors, and MIL aggregators, which increases the practical relevance of the findings. (4) The supervised variant, SAM-LP, is a sensible attempt to reduce signal dilution from irrelevant tissue regions.
- 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 ground-truth FE ranking is defined by averaging performance across multiple MIL aggregators, which may obscure important FE–aggregator interactions and does not fully reflect real-world model selection for a fixed downstream pipeline. (2) Although the reported rank correlations are promising, the improvements over baselines are not always large or consistent across datasets, and statistical uncertainty (e.g., variance, confidence intervals, significance tests) is missing. (3) The efficiency claim is insufficiently quantified: the paper argues for major computational savings, but does not provide concrete runtime, GPU-hour, or memory comparisons. (4) It remains unclear whether the gain comes specifically from SAM-based structural priors or more generally from any region-based grouping, since no ablation with simpler alternatives is provided. (5) The supervised protocol needs a clearer description, especially regarding label usage, subset construction, and separation from downstream evaluation to rule out potential leakage.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses a relevant and practical problem in computational pathology and presents a well-motivated, empirically promising solution. However, the current evidence is not yet fully convincing due to limited statistical analysis, insufficient efficiency quantification, and missing ablations on the role of SAM.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I think the idea is very interesting.
Author Feedback
We thank all reviewers for their constructive feedback and recognition of the problem’s practical importance (R1, R2, R4) and our extensive experimental setup (R1, R2, R4). We deeply value your insightful suggestions and address the main concerns below.
1.Computational efficiency quantification (R1, R2, R4) We will add a table reporting parameters, GPU memory, FLOPs, and wall-clock runtime for each component (FE, SAM, MIL aggregator, SAM-LP, SAM-Cluster), along with the following analysis. Exhaustive search costs: C_exhaustive = N × |Φ| × (K × C_FE + M × C_aggregator), where N is total WSIs, |Φ| candidate FEs, K average patches per WSI, M aggregators (9 in our study). Our framework costs: C_ours = n × [|Φ| × (K × C_FE + α) + C_SAM], using only n ≪ N WSIs and no MIL training. On Camelyon16 (RTX A6000, FE avg. 341M params, K = 4,090): K × C_FE ≈ 327 TFLOPs per WSI, while C_SAM ≈ 3.2 TFLOPs (<1% of single-FE extraction, run once on thumbnails) and α ≈ 3.7 MFLOPs (no backprop). Therefore C_ours / C_exhaustive ≈ n / N, with further savings from avoided aggregator training.
2.SAM vs. simpler regionalization alternatives (R1, R2, R4) We excluded simpler alternatives by design:
- Otsu: Our pipeline (Trident) already removes background, so Otsu cannot capture intra-tissue heterogeneity.
- Grid grouping: Ignores biological boundaries entirely.
- SLIC superpixels: Requires dataset-specific tuning (segment count, compactness), which contradicts our framework’s “zero-shot, out-of-the-box” philosophy — users would need to tune segmentation hyperparameters before they can even select an FE. However, we fully understand the necessity of quantitatively demonstrating that SAM’s regionalization outperforms other alternatives. We will address this point and include a detailed quantitative comparison in the subsequent journal extension.
3.When SAM-LP outperforms LP (R1, R2, R4) SAM-LP gains correlate with structural complexity (avg. patches and SAM regions per WSI):
- High complexity (SAM-LP > LP): TCGA-NSCLC (3,834 patches, 90.3 regions): ρ improves 0.6364 → 0.9515.BRACS (3,617 patches, 128.6 regions) shows similar gains.
- Low complexity (SAM-LP ≤ LP): PANDA is the smallest (200 patches, 19.7 regions), where LP/LogME suffices. TCGA-GLIOMA and UBC-OCEAN have many patches but limited structural separation (53–70 regions), yielding comparable performance. This provides actionable guidance: SAM-LP is most beneficial on large, structurally heterogeneous slides where signal dilution is severe.
4.Aggregator bias and ground-truth robustness (R1, R4) We averaged across 9 aggregators × 5 seeds to assess general FE transferability. To address the fixed-pipeline concern, our per-aggregator breakdown shows:
- The averaged GT ranking correlates strongly with individual aggregator rankings (mean ρ = 0.8205) and especially with the best-performing aggregator per dataset (ρ = 0.9299), confirming its representativeness.
- Per-aggregator stratification: SAM-LP correlates highest with Mean Pooling (0.6078) and RRT-MIL (0.5955), lowest with WiKG (0.5448), averaging 0.5688 ± 0.0226 across 9 aggregators. As R1 hypothesized, the linear-probing-based formulation aligns most closely with mean pooling. SAM-Cluster correlates highest with CLAM (0.5260) without a clear architectural pattern. Predictive power remains stable across pooling and attention-based aggregators.
5.Missing implementation details (R1, R4) Hardware specs, SAM configuration, and EM stopping criteria will be added to the manuscript and released code. Supervised label usage is documented in the code: patient-level train/val/test splits prevent leakage, and pre-evaluation metrics are computed only on training-set subsets (5 random samplings of n WSIs). All baseline results are averaged over 5 seeds. We will add variance to Table 1 and report significance tests for main results.
6.Other points (R2)
- Top-k: Fig. 4(c) reports Top-1 through Top-5.- Sample size: Fig. 4(a, b) analyzes n.
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 studies the practical problem of selecting WSI feature extractors before full MIL training and proposes a structure-aware pre-evaluation framework based on SAM-derived regional information. The reviewers find the problem practical and the empirical study reasonably broad. However, several concerns remain about the methodological justification, the consistency of the empirical findings, and the practical strength of the efficiency claim, etc. The authors are encouraged to address the concerns in the rebuttal.
- 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 rebuttal addresses most concerns such as clarifying the efficiency analysis, aggregator-specific behavior, dataset-dependent performance patterns, label usage, and implementation details. The revision should incorporate these clarifications and quantitative analyses into the manuscript.
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.
All three reviewers raised their scores to accept after rebuttal. The idea and the methodology novelty are recognized by the reviewers. The rebuttal addressed most of the concerns raised in the first round of review.
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.
clear accept for this paper.