Pith. sign in

REVIEW 5 major objections 6 minor 2 references

Fully Automatic Content-Aware Tiling Pipeline for Pathology Whole Slide Images

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A fully automatic pipeline can segment tissue folds, blur, background, and pen markings at pixel level and tile pathology slides with minimal tissue loss, reporting over 95% accuracy, precision, recall, and F1 across all artifact types.

desk verdict Useful preprocessing pipeline with a real generalization claim, but the paper's own external tables contradict the headline >95% metrics; worth reviewing after claims are recalibrated. read the letter →

arxiv 2411.16885 v2 pith:4M33QVIS submitted 2024-11-25 cs.MM

Falah Jabar , Lill-Tove Rasmussen Busund , Biagio Ricciuti , Masoud Tafavvoghi , Mette P{o}hl , Sigve Andersen , Tom Donnem , David J. Kwiatkowski
show 1 more author
Mehrdad Rakaee
This is my paper · ORCID
classification cs.MM
keywords wholeslideimagescomputationalpathologyartifactdetectionsemanticsegmentationcontent-awaretilingpen-markingremovaldeeplearninghistology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a single pixel-based segmentation model can detect the artifacts that corrupt pathology whole slide images, such as tissue folds, blurring, background, and pen markings, accurately enough to automate the tiling step every deep-learning pathology pipeline needs. On internal and external test sets, the model reports accuracy, precision, recall, and F1 above 95% across all artifact types, with Dice scores above 94%. The proposed WSI-SmartTiling pipeline then uses the segmentation to choose, from a center tile and its four overlapping neighbors, the tile with the fewest artifacts, recovering 0.76% to 7.5% more qualified tissue per slide than standard tiling. If these results hold on truly unseen slides, downstream models would get cleaner training data and retain more diagnostically relevant tissue.

What carries the argument

The load-bearing object is the artifact segmentation model: a dual-branch hierarchical global-local fusion network (built on the DHUnet design) with a Swin Transformer global encoder and a ConvNeXt local encoder, trained with pixel-level labels for background, qualified tissue, folds, and blur. Around it, the pipeline uses Otsu thresholding to find the tissue region, a color-percentage rule to filter heavy pen-marked tiles, a ResNet-18 plus Pix2Pix pen-removal stage, and a tile-selection cost $C=\lambda_{fo}P_{fo}+\lambda_{bl}P_{bl}+\lambda_{bg}P_{bg}$ with all weights set to 1, which selects the tile with minimal artifacts from each set of five overlapping candidates. The segmentation mask is what carries the claimed performance: it lets the pipeline exclude artifacts at pixel resolution rather than discarding whole tiles.

What would settle it

Re-train the same model on the same data but split at whole-slide level, so no slide contributes tiles to both training and test, and compare the internal test metrics: if accuracy, precision, recall, F1, or Dice fall well below the reported 95% and 94%, the original numbers were inflated by tile-level leakage. A second check is applying the pipeline to scanner formats it never trained on, such as NDPI or VSI, and measuring whether the metrics stay above 95%.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a dual-branch hierarchical global-local fusion network, with Swin Transformer for global context and ConvNeXt for local detail, trained at 20x and 40x with pixel-level labels for background, qualified tissue, fold, and blur, can segment pathology artifacts well enough to drive a fully automatic content-aware tiling pipeline. The model reaches over 95% accuracy, precision, recall, and F1 and Dice scores above 94% on the development test set and on an external 18-WSI, six-organ dataset, and it outperforms existing artifact detectors, which are tile-level classifiers or low-magnification segmenters. The pipeline adds color-based pen-marking detection, GAN-based pen-marking removal, and a cost-function tile selection step that chooses among a center tile and four 25%-overlapping neighbors the one with the lowest combined fold, blur, and background percentage. In a subjective test with two expert pathologists, the model's tile choice matched the experts in at least 8 of 10 tile sets.

Load-bearing premise

The reported internal accuracy assumes that splitting tiles randomly, instead of splitting whole slides, leaves the test set independent; since tiles from the same slide can end up in both training and testing, the high scores may partly come from slide-specific patterns rather than general artifact recognition.

Editorial extensions

If this is right

  • Downstream deep-learning models trained on WSI-SmartTiling output will see fewer fold, blur, and background tiles, which should reduce artifact-driven over- and under-detection in cell counting and classification.
  • Tissue previously lost near boundaries, in fragmented biopsies, or under pen marks can be recovered, with the paper measuring a 0.76% to 7.5% qualified-tissue gain per slide over standard tiling.
  • Because the artifact model works at 20x and 40x, it can feed the high-magnification tiles required by modern pathology foundation models without downsampling.
  • The pipeline can be inserted upstream of weakly supervised frameworks such as CLAM or STAMP, replacing generic tiling with artifact-aware tile selection.
  • Pen-marking detection emerges as a side effect of the artifact segmenter even though it was not trained for that task, offering a route to flag marked regions without a dedicated detector.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to weight the cost function per task: morphometry pipelines could penalize blur more heavily, while cell-counting pipelines could penalize folds more heavily, since all three weights are currently set to 1.
  • The artifact segmenter identifies pen-markings even though it was not trained on them; the same architecture could be tested on other marker-like structures, such as melanin granules in melanoma, which the authors mention as a future direction.
  • The 25% overlap search could be generalized to variable overlap or irregular tile placement, trading compute for even more tissue retention near irregular tissue borders.
  • Because the pipeline already supports 20x and 40x inputs, integrating it ahead of weakly supervised frameworks would give those frameworks high-magnification, artifact-free tiles without additional preprocessing development.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper describes WSI-SmartTiling, a fully automatic content-aware tiling pipeline for H&E whole slide images. The pipeline integrates a pixel-based semantic segmentation model that classifies regions into artifact-free tissue, fold, blur, and background; a pen-marking detection and removal step based on an existing GAN-based inpainting model; and a tile-selection optimization that minimizes artifacts and background while maximizing qualified tissue. The artifact detection model uses a dual-branch global-local fusion network combining Swin Transformer and ConvNeXt, trained on a lung-cancer dataset of approximately 25,000 tiles and validated on an external TCGA set of 18 WSIs from six organs. The paper claims consistently above 95% accuracy, precision, recall, and F1 across artifact types, and above 94% Dice, outperforming existing state-of-the-art methods such as GrandQC and MoE-CNN. The pipeline's tiling is evaluated against expert pathologist selections on 100 tile sets from 10 WSIs, with reported agreement of at least 80% and qualified-tissue gains up to 7.5% over standard tiling.

Significance. The contribution is potentially valuable: the pipeline addresses a real preprocessing need in computational pathology, the dataset is moderately large and includes two scanners, and the external validation on TCGA is a genuine effort at generalization testing. The authors also provide code access (though without a linked repository in the manuscript) and the integration of pen-mark removal with artifact-aware tiling is a useful engineering contribution. However, the central quantitative claim of >95% across all artifact types is contradicted by the paper's own supplementary confusion matrices, and the internal evaluation split is at tile level, which risks inflating metrics. If these issues are resolved and claims are corrected, the pipeline could still be a solid contribution, but the current evidence does not support the headline numbers.

major comments (5)
  1. [Section 4.2 / Table S2] The claim that the model 'consistently achieved over 95% accuracy, precision, recall, and F1 score across all artifact types' is not supported by the external validation confusion matrices. For Breast, fold recall is 168/(66+2+168)=71.2% and F1 is 83.2%; for Kidney, fold recall is 142/(37+1+142)=78.9%. Even pooled across all six organs, fold recall is 1232/(110+8+1232)=91.3%, below the stated 95%. The supplementary text to Fig. S4 restates the >95% claim, which is directly contradicted. Please revise the abstract, Section 4.2, and Fig. S4 caption to report the actual per-class metrics, and provide a corrected version of the strong generalization claim.
  2. [Section 3.2.4] The development dataset is randomly split at the tile level (87% training / 13% testing). Since multiple tiles originate from the same WSI, this allows information leakage between training and test sets, artificially inflating the internal metrics in Table S1. A WSI-level split should be used, or the internal metrics should be re-reported with a WSI-exclusive split to demonstrate that the >95% internal performance is not an artifact of tile leakage.
  3. [Section 4.1] The benchmark methods GrandQC, MoE-CNN, MoE-ViT, multiclass-CNN, and multiclass-ViT are listed, but the manuscript does not state whether they were retrained on the development dataset. If they were used as pre-trained models without adaptation, the comparison is not apples-to-apples and the claimed superiority may reflect mismatch between their training data and the evaluation data. Please clarify the training status of each benchmark; if they were not retrained, either retrain them under the same protocol or moderate the comparative claims.
  4. [Tables S1/S2 and Fig. 3] The artifact segmentation model has four output classes (background, qualified tissue, fold, blur), but the confusion matrices and classification metrics report only three classes (artifact-free, blur, fold). Background is a crucial class for tiling decisions (Section 3.2.5 computes background percentage), and its exclusion means the reported 'accuracy across all artifact types' does not cover the background class. Please report per-class metrics for background as well, or explicitly justify its exclusion from evaluation.
  5. [Section 3.2.2] The pen-marking removal step is presented as a contribution (Introduction item 3 and Section 3.2.2), but its output is never quantitatively evaluated in this study. The authors rely on the prior work of Ramanathan et al., yet the integration into WSI-SmartTiling could affect downstream tile selection. Please provide a quantitative or at least a systematic qualitative evaluation of pen-marking removal on a sample of the development or external WSIs.
minor comments (6)
  1. [Section 3.1.1] The total tile count is inconsistent: '24,942 tiles' is the sum of 9,828 + 7,334 + 7,780, but the sentence then adds 'Additionally, 1000 background tiles,' making the total 25,942. Clarify the total and indicate how the background tiles are used in training/testing.
  2. [Figures S5 and S6] The captions contain the typo 'QrandQC' instead of 'GrandQC'.
  3. [Fig. S4 caption] The caption of Fig. S4 repeats the >95% claim that is contradicted by Table S2; this caption should be corrected alongside the main text (see Major Comment 1).
  4. [Section 3.3 and Fig. 6] The description of the expert agreement test says the model achieved 'at least 80% accuracy,' but the method states 10 sets per WSI; clarify whether the 80% refers to a threshold per WSI or an overall average, and report confidence intervals given the small sample (10 WSIs).
  5. [Eq. (1) and Section 3.2.5] The weighting parameters (lambda_fo, lambda_bl, lambda_bg) are all set to 1, and P_min/P_max in Section 3.2.2 are set by hand; a sensitivity analysis of these hand-set parameters would strengthen the pipeline's robustness claims.
  6. [Data/code availability] The Data/code availability statement says 'code is available on GitHub' without providing a repository URL or DOI; please provide a persistent link or accession identifier.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the artifact-detection and tiling results are benchmarked against manual annotations and external data, not derived from the model's own outputs.

full rationale

The paper's central claims are empirical performance evaluations of a supervised segmentation model against manually annotated ground truth, internal test tiles, and external TCGA WSIs. The artifact segmentation model is trained on labeled tiles (Section 3.1.3) and evaluated with independent confusion matrices and Dice scores; the reported metrics are not constructed from the model's own outputs. The pen-marking detection thresholds and cost-function weights in Eq. (1) are hand-set hyperparameters, not fitted parameters renamed as predictions. Self-citations to Rakaee et al. (2023a, 2023b) are used for data provenance (source WSI cohorts), not to prove the model's performance or to forbid alternative methods. The architecture is based on prior work by Wang et al. (2023), and the pen-removal GAN is cited from Ramanathan et al. (2023); both are external references used as building blocks, not as a self-citation chain that forces the result. The most serious issue in the paper is an internal-consistency problem, not circularity: the abstract and main text claim metrics 'above 95% across all artifact types,' while the supplementary Table S2 confusion matrices appear to imply lower fold recall on some external organs (e.g., breast fold recall 168/(66+2+168) ≈ 71% and kidney fold recall 142/(37+1+142) ≈ 79%). That contradiction is a correctness and reporting concern, not a circular step. Likewise, the tile-level random split described in Section 3.2.4 may cause train/test leakage and inflated internal metrics, but that is an evaluation-validity concern rather than a derivation that reduces to its inputs. No equation is defined in terms of the quantity it predicts, no fitted value is relabeled as a prediction, and no load-bearing premise rests solely on a self-citation. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on supervised labels and a small external validation set; no new physical or model entities are introduced. The main hand-set parameters are cost weights, pen thresholds, and tile overlap, which affect tiling behavior but not the segmentation metrics directly.

free parameters (3)
  • Cost function weights lambda_fo, lambda_bl, lambda_bg = 1, 1, 1
    All weighting parameters in Eq. (1) are set to 1 without sensitivity analysis; the tile-selection result may change if weights are tuned.
  • Pen-marking thresholds P_min and P_max = P_min = 0.2, P_max = 0.9
    Hand-chosen thresholds in Section 3.2.2 for classifying tiles by pen-marking level; no optimization or robustness test is reported.
  • Tile overlap parameter O_t = 25%
    Chosen overlap for neighboring tiles in Section 3.2.1; not varied across experiments.
assumptions (3)
  • domain assumption Manual annotations of folds and blur plus HSL thresholding of unannotated regions yield correct pixel-level ground truth labels.
    The supervised model is trained and evaluated against these labels; errors in annotation, especially around ambiguous tissue, would be learned as ground truth. See Section 3.1.3.
  • domain assumption The four classes (background, qualified tissue, fold, blur) are sufficient and mutually exclusive for WSI quality assessment.
    The model cannot represent other artifacts such as pen markings, air bubbles, or staining variations; the authors later note melanophages could be misclassified. See Sections 3.2.3 and 5.
  • domain assumption External TCGA WSIs (3 per organ) are representative of the six organs and staining variations.
    Generalization across organs is claimed from a small external set; the sample is thin for organ-level conclusions. See Section 3.1.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fully Automatic Content-Aware Tiling Pipeline for Pathology Whole Slide Images." pith.science (2026). https://pith.science/paper/4M33QVIS

@misc{pith2026241116885,
  author       = {Pith},
  title        = {Pith review of: Fully Automatic Content-Aware Tiling Pipeline for Pathology Whole Slide Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4M33QVIS}},
  note         = {Machine review of arXiv:2411.16885}
}
read the original abstract

In recent years, the use of deep learning (DL) methods, including convolutional neural networks (CNNs) and vision transformers (ViTs), has significantly advanced computational pathology, enhancing both diagnostic accuracy and efficiency. Hematoxylin and Eosin (H&E) Whole Slide Images (WSI) plays a crucial role by providing detailed tissue samples for the analysis and training of DL models. However, WSIs often contain regions with artifacts such as tissue folds, blurring, as well as non-tissue regions (background), which can negatively impact DL model performance. These artifacts are diagnostically irrelevant and can lead to inaccurate results. This paper proposes a fully automatic supervised DL pipeline for WSI Quality Assessment (WSI-QA) that uses a fused model combining CNNs and ViTs to detect and exclude WSI regions with artifacts, ensuring that only qualified WSI regions are used to build DL-based computational pathology applications. The proposed pipeline employs a pixel-based segmentation model to classify WSI regions as either qualified or non-qualified based on the presence of artifacts. The proposed model was trained on a large and diverse dataset and validated with internal and external data from various human organs, scanners, and H&E staining procedures. Quantitative and qualitative evaluations demonstrate the superiority of the proposed model, which outperforms state-of-the-art methods in WSI artifact detection. The proposed model consistently achieved over 95% accuracy, precision, recall, and F1 score across all artifact types. Furthermore, the WSI-QA pipeline shows strong generalization across different tissue types and scanning conditions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

  1. [1]

    Berman, A.G. et al. (2021) ‘PathML: A Unified Framework for Whole -Slide Image Analysis with Deep Learning’, MedRxiv Preprnt, pp. 1–1. Available at: https://doi.org/10.1101/2021.07.07.21260138 Haghighat, M. et al. (2021) ‘PathProfiler: Automated Quality Assessment of Retrospective Histopathology Whole-Slide Image Cohorts by Artificial Intelligence – A Cas...

    work page Pith review arXiv 2021

  2. [120]

    Available at: https://doi.org/10.1038/s43856-022-00186-5 Weng, Z. et al . (2024) ‘GrandQC: A Comprehensive Solution to Quality Control Problem in Digital Pathology’, Nature Communications, 15(1), p. 10685. Available at: https://doi.org/10.1038/s41467-024-54769-y

    doi:10.1038/s43856-022-00186-5 2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.