List of Papers Browse by Subject Areas Author List
Abstract
To ensure the highest quality in cancer research and computer aided diagnostics in computational pathology, Whole Slide Image (WSI) quality should be carefully monitored. While several open-source Quality Control (QC) tools exist, their comparative performance remains largely unaudited. In this work, we present the first independent, cross-species benchmark for WSI QC solutions focused on the critical task of automated artifact detection. Recognizing that model generalization is often hindered by narrow training data, we curated a diverse artifact-centric dataset featuring 11 distinct artifact types across 6 tissue types, 2 species, and 4 independent data sources. We evaluate both pixel-level segmentation and patch-level classification methods, normalizing their outputs and comparing them to expert-annotated WSIs. Performance is quantified using Dice score, accuracy, precision, and recall within tissue containing region. Our results reveal a substantial performance gap; average artifact detection Dice scores do not exceed 0.6, with significant performance variability across dataset subcategories. To support ongoing development, we release our dataset at https://doi.org/10.5281/zenodo.20339462 and our benchmarking tool at https://www.codabench.org/competitions/16457.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3442_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://www.codabench.org/competitions/16457/
Link to the Dataset(s)
https://doi.org/10.5281/zenodo.20339462
BibTex
@InProceedings{KacGab_SlideGuard_MICCAI2026,
author = { Kaczmarek, Gabriela AND Krawczyk-Borysiak, Zuzanna AND Miller, Mateusz AND Krawczyk, Adam AND Sokol, Malgorzata AND Przybylska, Martyna AND Szymanski, Lukasz AND Markiewicz, Tomasz AND Swiderska-Chadaj, Zaneta},
title = { { SlideGuard: WSI Artifact Detection Benchmark } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces a novel H&E-stained WSI dataset designed to investigate the performance of artifact detection methods, benchmarks several existing methods, and provides a comprehensive analysis of the results.
- 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 is very well written and easy to follow. Its organization is sound.
The authors designed the first (soon to-be) publicly available dataset for artifact detection, filling an important gap in the highly unstandardized digital pathology ecosystem. Furthermore, the dataset covers multiple organs (and indications) and species, with data originating from three sources for human WSIs and one source for mouse WSIs. This enables the benchmarking of artifact detection methods in a non-trivial and realistic setting. In addition to slides containing artifacts, the authors also included near-perfect slides, as negative controls.
The benchmark itself is convincing; the authors evaluate the main relevant methods, and account for the many discrepancies between existing methods (e. g. pixel-level segmentation of artifacts versus patch-level classification, or inconsistencies in artifact-type labels). The chosen metrics are relevant, and the results section includes comments on the limitations of these metrics in the context of the benchmark.
The discussion section is thorough, suggests guidelines to drive the community forward and addresses the limitations of the work (e. g. partial annotation subjectivity).
- 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.
While the paper does not suffer from major weaknesses, we believe it could benefit from the following clarifications:
-
The central claim that artifacts « impair the performance of automated algorithms » is currently backed by a single citation from 2021.Given the significant progress of computational pathology in the last 5 years (e. g. the paper cited by the authors does not use any pathology foundation models, which are now the backbone of most modern methods), it would be valuable to include more recent studies. Recent artifact detection methods, for instance, may provide more information on this topic.
-
How the Dice score and other metrics are computed across settings is partially unclear.
-
For instance, the authors state « Artifact detection performance was quantified using class-specific Dice scores, accuracy, precision, and recall. » but accuracy is not computed, and the following sentence can be found slightly below « Due to divergent artifact taxonomies among the benchmarked methods, class-specific precision and Dice score calculations were infeasible. ». This discrepancy likely stems from the fact that « class » equates to artifact type in the second statement, and « class » equates to artifact or not in the first, but this is initially confusing for the reader. Likewise, the subsequent sentence uses the term « label-agnostic recall » to indicate the recall for the artifact class. Standardizing the nomenclature would make the paper easier to read.
-
The authors state « To compare methods producing artifact masks (pixel-level) and those calculating per-patch predictions, ground truth was discretized to each patch-level method’s native grid; a patch was labeled as an artifact if annotation coverage exceeded 25%. » Does this mean that segmentation methods are evaluated with a Dice score computed pixel-wise over the full WSI (or is there a per-patch computation averaged across the slide ?) and that the Dice is computed patch-wise across the full WSI for patch-level classification methods ? Likewise, how was the 25% cut-off chosen ?
- If possible, given the constrained MICCAI format and available information, it would be interesting to include the magnification used for all benchmarked methods and the sets of classes in the comparison table (Table 1). Additionally (again, if possible), it would be helpful to indicate explicitly the training sources of each of the methods (as opposed to only providing the number) as to enable the detection of a possible over-fitting to the training set artifacts. This would also better contextualize the discussion subsection around the fuzzy mapping between nominal magnification and resolution in MPP.
- The authors state « High recall for artifact categories not included in a model’s native label set indicates either cross-class false positives or a generalized sensitivity to morphological deviations from artifact-free tissue. ». Having the sets of classes for each method explicitly stated or via an asterisk added to the per-artifact recall table for instance (for the cases where the artifact type does not exist for a given method) would enable the reader to understand the situations in this actually happens compared to situations where the method is successfully able to point-out the specific artifact type.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
I recommend an Accept for this submission.
The authors present a valuable contribution to the computational pathology community by introducing the first publicly available dataset for WSI artifact detection. The dataset covers a comprehensive range of tissue types and centers across two species, providing a realistic setting for QA methods. The subsequent benchmarking study is rigorous.
Although there are a few minor points that could benefit from clarifications, the paper does not suffer from any major weaknesses.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
In this paper, the authors curated a novel benchmark dataset focused on the task of automated WSI artifact detection, comprising of 6 tissue types, 2 species, and 4 independent data sources. This data also contains pixel-level annotations of the artifacts. Additionally, they also evaluated pixel-level segmentation, and patch-level classification with this dataset and compared them to expert annotated WSIs.
- 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.
Motivation: This paper addresses a major challenge in the field of digital pathology. Artifacts are often found in WSIs and there are several automated QC methods (histoQC, pathprofiler, DKL, TCNN, etc.), each with its own limitations. Such a curated dataset would aid in the development of robust artifact-detection and quantification algorithms. Rich dataset: Multi-source, multi-species dataset encompassing 11 artifact categories across 6 tissue types, 2 species, 6 organs, and 4 datasets. Rich pixel-level annotations: 2,116 detailed pixel level annotations for the eleven artifact classes Table 3 highlights variations in performance of different automated QC methods, warranting the need for such a dataset that might help build better QC 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.
o Fig 1 is unclear. The ground truth artifact mask shows the tissue as foreground and the rest as background. In this example, is the entire foreground an artifact? Or does the black region in the center of the mask represent the artifact? o The annotations were performed by a single reader. This may add subjectivity and bias to the dataset. o Although the dataset is valuable, the algorithms developed for evaluation i.e., to segment the tissue or to label the patches we not optimal. This may affect the results: o The tissue detection steps involve basic methods like “intensity thresholding at 200”, which is not a robust approach and the performance can vary with respect to staining variations. o “a patch was labeled as an artifact if annotation coverage exceeded 25%”. The algorithm outputs are at patch level although the annotations are at pixel level leading to loss of information.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
In this paper, the authors curated a novel benchmark dataset focused on the task of automated WSI artifact detection. This seems to be an extensive dataset that may be useful for the digital pathology community to develop stronger, more robust automated QC methods. The results in the paper show that the performance of existing QC methods vary quite a bit compared to manual annotations, thereby highlighting the need to build a method that’s more robust and this dataset may be helpful.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
In this paper the authors present a third-party independent benchmark dataset and pipeline to evaluate existing WSI Quality Control methods for artifact detection. Their experiments and comparison span across 5 segmentation models and 2 patch classifiers (7 distinct methods), evaluated on 53 H&E WSIs. With these comparisons, the authors provide insights into current limitations of QC solutions and establish a standardized clinical evaluation framework to help the community developing deployable and robust models for artifact detection.
- 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) First Independent third-party benchmark dataset for WSI artifact detection, from multiple sources (4 centers), species (human/mice), tissue types (6), and broad artifact taxonomy (11 artifacts). While annotated from a single expert, this is in line with practices of previous benchmarks for WSI QC quantification (GrandQC, Weng et al., Nature Communications 2024).
2) Breadth of method comparison, with evaluation spanning 7 methods from 2019 to 2024, 5 segmentation models and 2 patch classifiers (TCNN and PathProfiler), with these methods trained on varying data scales and domains. This is a valuable comparison for the community on existing tools, showcasing out-of-the box performance on a real world dataset.
3) Release of a practical standardized benchmarking tool for future QC method evaluation, promising to lower the barrier for future developers to situate their work within a common framework.
4) Importance of findings, that while methods reliably preserve artifact-free tissue, they do not perform consistently in artifact localization tasks across different sources and taxonomies.
- 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 first benchmark to test WSI QC methods” claim is overstated and misleading. GrandQC (Weng et al.) already provides a publicly released annotated test dataset of 318 WSIs across 4 organ types with pixel-level ground truth and a head-to-head comparison of multiple QC tools, making it a de facto benchmark for this task. SlideGuard’s genuine and significant novelty is that it is the first independent benchmark for this task, with a broader diversity of samples (multi-species, multi-center, 11 artifact classes). The authors need to reframe this contribution precisely along these lines, independence and diversity versus scale, and additionally verify and disclose whether any overlap exists between SlideGuard’s TCGA slides and GrandQC’s training set, as this would directly affect the validity of GrandQC’s evaluation within this benchmark.
2) Annotator subjectivity described in the ‘Detection Asymmetry’ paragraph and Conclusions section undermines the reliability of the dataset’s ground truth. This statement directly contradicts the single-annotator protocol: if boundary subjectivity is significant enough to alter or explain systematic performance gaps across all evaluated methods, then the absence of an extra annotator and inter-rater agreement analysis is a critical methodological gap, not a mere limitation. This is especially consequential for a benchmark paper, whose primary contribution is the ground truth itself. While GrandQC also relied on a primary single annotator for the test set, authors explicitly argue on the peer review process file, that inter-rater agreement is less critical for artifact detection due to its relatively objective nature. SlideGuard takes the opposite position by acknowledging annotation subjectivity as a key driver of performance variability, making the absence of inter-rater analysis a self-inflicted and particularly consequential contradiction that the authors must address.
3) At 53 WSIs, the dataset is substantially smaller than comparable benchmarks (GrandQC test set: 318 WSIs). No per-class annotation statistics are reported, making it impossible to assess class imbalance or evaluate whether rare artifact categories are sufficiently represented (should be reported in both table 2 and 3). Given this limited scale, the absence of uncertainty estimates, no standard deviations, confidence intervals, or significance tests, makes it difficult to quantify how different performance between methods is.
4) The inference resolution is never specified: the MPP or nominal magnification at which each method processes WSIs is not reported or standardized across the benchmark, despite the authors themselves identifying MPP inconsistency as a key practical limitation in the Discussion. This is a direct internal contradiction that must be resolved. The authors should explicitly state the extraction MPP for each method, clarify whether this is fixed or method-dependent, and discuss how resolution differences across methods may confound the reported performance comparisons. Furthermore, the recommendation to shift from nominal magnification to MPP in the Discussion is presented as a practical requirement but is entirely unsupported by the benchmark itself, no ablation or sensitivity analysis is performed comparing method performance across different MPP values (under the same nominal magnification), which would be the natural way to empirically ground this claim.
5) How sensitive is the performance of patch classifiers to the discretization threshold? An ablation on 10%, 15%, 20%, 25%, 50% and 75% might explain how reliable are current results for TCNN and PathProfiler. The choice on 25% is not empirically validated.
6) Table 2 reports Dice scores for all methods without distinguishing between pixel-level spatial overlap Dice for segmentation methods and patch-level F1 score for TCNN and PathProfiler, which is what Dice mathematically reduces to when a single binary label is assigned per patch. These quantities should be clearly and distinctly presented. The authors must explicitly state what metric is being reported for each method type and clearly flag the weaknesses of cross-paradigm comparison.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
Minor weaknesses: 1) Table 1 has column alignment issues; namely, the ‘Backbone’ and ‘Dataset Characteristics’ columns appear not centered, making this table visually aesthetically weird. 2) Table 3 has some results bold without being the best for example ‘Precipitates’ and ‘Ink’ performance of Path profiler. In this table it would be useful to have the actual dataset percentage next to the artifact to show how meaningful current comparison are based on annotated size.
- 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 provides the community with a valuable benchmark dataset for QC evaluation. However, current score is justified by four distinct issues: the novelty claim requires reframing as GrandQC already provides a public benchmark; the paper self-contradicts by invoking annotation subjectivity to explain performance gaps which, if significant, requires a second annotator and inter-rater agreement analysis; critical dataset characterization is missing including per category annotation distributions, uncertainty estimates in metrics, and inference resolution; and the MPP recommendation in the discussion section is unsupported by sensitivity analysis.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers for their thoughtful and constructive feedback. We are encouraged by the reviewers’ positive feedback on SlideGuard’s relevance and have carefully considered each suggestion to improve the manuscript. Below, we address the technical and editorial points raised.
Regarding annotator subjectivity: We distinguish between semantic agreement and geometric precision. Artifact classification is an objective task for experienced histotechnicians. However, for artifacts like dust (soft edges) or ink (complex structures), manual polygons lack the pixel-level granularity of machine segmentation. This inherent spatial ambiguity, rather than annotator bias, may limit attainable Dice scores.
Addressing magnification used by models: To ensure a fair comparison, models were evaluated using their native resolution requirements. Forcing an incorrect scale would inherently compromise the benchmark. Patch-level methods were provided with their exact requested MPP, whereas pixel-level methods processed the WSIs directly to manage resolution internally. Running tools ‘off-the-shelf’ this way exposes how they handle physical scale inconsistencies across centers, providing a realistic sensitivity analysis.
In regard to the comparison modes: Pixel-level methods were evaluated pixel-wise and patch-level methods were evaluated patch-wise using their native output grids. The 25% discretization threshold was chosen empirically based on previous works, and we acknowledge the importance of a sensitivity study in future works. We also acknowledge that patch-level methods benefit from this task simplification, and we explicitly flag the weaknesses of this cross-paradigm comparison in the Results. To further address this, Tables 2 and 3 have been updated in the camera-ready version to group methods by output type to clearly distinguish these evaluation modes.
As to the intensity masks: They were manually verified to ensure accurate ROIs. This step solely excludes background, focusing analysis on artifact localization rather than the models’ internal tissue detection. Large-scale datasets in the future will adopt more robust segmentation. We are thankful to the reviewer for pointing it out as a future direction.
To address the “first benchmark” phrasing: We have reformulated the claim to more accurately reflect SlideGuard’s contribution as the first independent, cross-species benchmark. SlideGuard’s value lies in its high taxonomic granularity (11 artifact classes) and its inclusion of non-TCGA human and animal samples, providing a guarantee of unseen data samples.
As to the expansion of tables: Thank you for the constructive suggestions to expand the comparison in Table 1 (including magnification and artifact types) and to provide a broader analysis of the annotation dataset. However, due to strict page constraints and the inability to include supplementary material, we are unable to accommodate these additions. Furthermore, presenting raw annotation metrics or dataset percentage alone, as a mere numerical addition to Table 3, would be misleading without a thorough contextual discussion of the variations in artifact sizes that we unfortunately cannot fit within the current space limits. Relevant information will, however, be included on the Grand Challenge page of the project.
Regarding presentation and editorial clarity: We have expanded the literature review on the downstream impact of artifacts, unified our terminology (consistently using ‘class’, ‘label’, and ‘category’), and improved mask visibility in Figure 1.In Table 3, we also identified the bolding errors post-submission and have corrected the table along with the relevant commentary in the Results to accurately reflect the model rankings. These refinements, along with all other formatting suggestions, will be incorporated into the camera-ready version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
SlideGuard proposes a heterogeneous multi-source artifact dataset and harmonized evaluation of segmentation- vs patch-QC detectors. The reviewers appreciated the clarity, community needs for standardized quality control benchmarking, and richness of taxonomy/sources. Conscientious limitation discussion on metrics, and partial subjectivity (R2–4). Raised issues include sharpening evidence that artifacts degrade modern pipelines versus a single dated citation plus clarifying Dice/accuracy nomenclature and patch-threshold derivation (R2), overstated “first benchmark” claim versus GrandQC and questions on overlap/train contamination (R4), interplay of annotation subjectivity with single-reader protocol lacking inter-rater study (R4), comparatively small corpus vs some benchmarks lacking uncertainty/per-class prevalence detail (R4), and presentation issues in comparative tables noted as minor polish (R4). The AC read the paper, reviewed it, and agrees that the paper has merit and that the benchmark with the dataset would be of interest to the community. The authors should address the raised concerns in their revised version.