{"id":"1586f7ce-e31f-49c6-b691-608f7865244b","arxiv_id":"2411.16885","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"WSI-SmartTiling automatically segments artifacts and background in whole slide images and tiles only qualified tissue, reporting high accuracy on internal and external datasets.","lead":"This paper describes WSI-SmartTiling, an automated pipeline that detects artifacts such as folds, blur, and background in pathology whole slide images and tiles only the qualified tissue regions. It combines a pixel-level segmentation model with pen-mark removal and a tile-selection step, reporting over 95% accuracy on internal and external test sets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own external-validation confusion matrices (Table S2) contradict the central >95% claim: fold recall is ~71% for breast and ~79% for kidney.","rationale":"The reader's weakest assumption was that tile-level rather than WSI-level splitting could inflate internal metrics. That is a legitimate methodological concern, but it is not the most load-bearing issue. The paper's own supplementary Table S2 contains the external-validation confusion matrices for the proposed model, and simple arithmetic from those matrices shows that the headline claim of '>95% accuracy, precision, recall, and F1 across all artifact types' is not met: external fold recall is 91.3% pooled, and only 71.2% for breast and 78.9% for kidney. The Fig. S4 statement that all metrics exceed 95% across all organs is contradicted by the very table it accompanies. This is an internal inconsistency, so no external benchmark or consensus is needed to expose it. It directly undermines the strongest claim in the abstract and Section 4.2, and it is not merely an 'addressable weakness' of experimental design; it is a false quantitative statement in the current manuscript. A conditional acceptance could be appropriate after the authors correct the overclaims and re-audit all reported metrics, but as submitted the central claim is not supported by the paper's own data. Therefore the recommended verdict is REJECT rather than UNCHANGED or CONDITIONAL.","tokens_in":17071,"tokens_out":6001,"duration_ms":56137,"concrete_test":"Recompute per-class recall, precision, and F1 for the proposed model from Table S2 (external validation), focusing on the fold row for Breast and Kidney and the pooled fold row across all organs. If the values reproduce (Breast recall ~71%, Kidney recall ~79%, pooled recall ~91%), the >95%-across-all-artifact-types claim is false as written; the manuscript must then either restrict the claim to the development test set or correct the abstract, Section 4.2, and Fig. S4 to report the actual per-class and per-organ ranges.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim, stated in the abstract and Section 4.2, is that the artifact-detection model 'consistently achieved over 95% accuracy, precision, recall, and F1 score across all artifact types' and generalizes strongly. The paper's own supplementary Table S2 provides per-organ confusion matrices for the proposed model on the external TCGA set. Recomputing from those matrices: for Breast, fold recall is 168/(66+2+168) = 71.2%, and F1 is 2*168/(168+236) = 83.2%; for Kidney, fold recall is 142/(37+1+142) = 78.9%. Even pooling all six organs, the external fold recall is 1232/(110+8+1232) = 91.3%, below the claimed 95%. The supplementary text for Fig. S4 explicitly states that the model 'consistently achieves performance for all metrics above 95% across all organs,' which is directly contradicted by Table S2. This is an internal inconsistency in the paper's own evidence, not a disagreement with external consensus. The tile-level random split issue identified by the reader is a real secondary concern, but the reported metrics fail the stated threshold even before considering possible train/test leakage.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":17321,"tokens_out":6932,"duration_ms":53571,"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":[{"comment":"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.","section":"Section 4.2 / Table S2"},{"comment":"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.","section":"Section 3.2.4"},{"comment":"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.","section":"Section 4.1"},{"comment":"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.","section":"Tables S1/S2 and Fig. 3"},{"comment":"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.","section":"Section 3.2.2"}],"minor_comments":[{"comment":"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.","section":"Section 3.1.1"},{"comment":"The captions contain the typo 'QrandQC' instead of 'GrandQC'.","section":"Figures S5 and S6"},{"comment":"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).","section":"Fig. S4 caption"},{"comment":"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).","section":"Section 3.3 and Fig. 6"},{"comment":"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.","section":"Eq. (1) and Section 3.2.5"},{"comment":"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.","section":"Data/code availability"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between the >95% claim and the supplementary confusion matrices (Table S2) is the most serious issue; the authors must correct both the claim and the supplementary caption before the paper can be considered further. The external validation is also very small (18 WSIs, 3 per organ), so even after correcting the metrics, the authors should temper generalization claims and acknowledge the per-organ sample-size limitation. The tile-level split concern is well-founded and should be addressed with a WSI-level split or explicit justification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: WSI-SmartTiling is a solid engineering contribution. It combines a known dual-branch segmentation network, a pen-mark removal GAN, and a cost-based tile selection rule into one pipeline that runs at 20x/40x, and the authors validate on an external TCGA set across six organs. That integration, plus the large annotated tile set, is genuinely new and useful for computational pathology labs that need artifact-free tiles. The tissue-gain numbers (0.76–7.5%) are modest but plausible, and the pathologist agreement test is a nice touch.\n\nThe soft spots are real, and one is load-bearing. The abstract and Section 4.2 claim accuracy, precision, recall, and F1 above 95% across artifact types. The paper's own supplementary Table S2 does not support that on the external set: for breast, fold recall is 71.2%; for kidney, 78.9%; pooled external fold recall is 91.3%. The supplementary text for Fig. S4 repeats the >95% claim, so this is an internal inconsistency, not an outside critique. That needs to be fixed before the paper is publishable as-is.\n\nThe second issue is the internal split. Section 3.2.4 says the development set was split randomly at tile level. Tiles from the same WSI can appear in both training and test, which can inflate the reported internal numbers. A WSI-level split is the right evaluation for a model meant to generalize to new slides. The external TCGA set is genuinely independent but small (18 WSIs, 3 per organ), so the generalization claim is thinner than the prose suggests.\n\nMinor points: the benchmark models are not described as retrained on the development data, so the comparison may be unfair; pen-mark removal is never quantitatively evaluated; and some table entries are inconsistent with the text. None of these are fatal. The core idea is sound, and the code is promised on GitHub.\n\nWho this is for: labs building preprocessing pipelines for H&E WSIs who want a ready-made artifact segmentation and tiling tool. It deserves a serious referee. I would send it out with a request for major revision, primarily to fix the metric claims, redo the internal split, and clarify benchmark training. The pipeline is worth engaging with, but the paper's current numbers overstate what the evidence shows.","headline":"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.","tokens_in":17867,"tokens_out":2057,"would_cite":false,"duration_ms":17832,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["whole slide images","computational pathology","artifact detection","semantic segmentation","content-aware tiling","pen-marking removal","deep learning","histology"],"falsifier":"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%.","tokens_in":16899,"feed_emoji":"🔬","tokens_out":9734,"duration_ms":79493,"temperature":0.7,"pith_summary":"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.","feed_headline":"Smart tiling pipeline cuts pathology slide artifacts with 95% accuracy","feed_subtitle":"Pixel-level model flags folds, blur, and pen marks so deep-learning pathology models train on clean tissue tiles.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the DHUnet dual-branch global-local fusion architecture on which the artifact segmentation model is built.","marker":"Wang et al., 2023"},{"why":"Supplies the ResNet-18 plus Pix2Pix pen-marking detection and removal pipeline used for cleaning tiles.","marker":"Ramanathan et al., 2023"},{"why":"Defines the MoE-CNN and MoE-ViT tile-wise artifact classifiers that serve as the main benchmarks.","marker":"Kanwal et al., 2024b"},{"why":"GrandQC is the pixel-wise artifact segmentation baseline the proposed model is compared against on Dice and qualitative outputs.","marker":"Weng et al., 2024"},{"why":"Provides the DFCI NSCLC WSI cohort that forms part of the development dataset.","marker":"Rakaee et al., 2023a"},{"why":"Provides the TNM-I NSCLC WSI cohort that forms the other part of the development dataset.","marker":"Rakaee et al., 2023b"},{"why":"Supplies the external TCGA dataset with 18 WSIs across six organs used to test generalization beyond the training cohorts.","marker":"National Cancer Institute, 2024"}],"fun_headline_variants":["Automatic pipeline removes pathology slide artifacts with 95% accuracy","CNN-ViT fusion model achieves 95% accuracy on pathology artifacts","Fully automatic tiling pipeline skips artifacts in whole slide images","AI tiling cleans pathology slides to boost deep learning accuracy","Artifact-free tiles from pathology WSIs via CNN-Swin fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Automatic pipeline removes pathology slide artifacts with 95% accuracy","CNN-ViT fusion model achieves 95% accuracy on pathology artifacts","Fully automatic tiling pipeline skips artifacts in whole slide images","AI tiling cleans pathology slides to boost deep learning accuracy","Artifact-free tiles from pathology WSIs via CNN-Swin fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000483,"raw_usage":{"total_tokens":2433,"prompt_tokens":1041,"completion_tokens":1392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":1303}},"tokens_in":657,"tokens_out":1392,"duration_ms":9756,"temperature":1.0,"reasoning_tokens":1303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:46:04.424999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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%.","supporting_citations":[],"review_version":1}