Slide quality control: six artifact-segmentation tools compared on three datasets, HistoKit ahead
A team from the Silesian University of Technology (Poland) and LMU Klinikum in Munich compared six public tools that flag preparation or scanning defects (folds, blur, air bubbles, stains) on digitised pathology slides. HistoKit, which reuses GrandQC's artifact module, comes first and HistoQC last, and a simple post-processing step improves HistoKit further. The work is useful because it finally compares these tools on the same data, but the annotated sets are tiny, one of them was annotated by a single person, and no result is translated into an effect on diagnosis.
The context
Digital pathology means scanning glass slides into giant images, "whole slide images" (WSI), several gigabytes each, so that computers can analyse them. Slides stained with hematoxylin and eosin (H&E, the standard stain that colours nuclei blue-purple and the rest pink) are fragile: a tissue fold, an air bubble under the coverslip, a blurred area, a marker-ink stain or a speck of dust can mislead a downstream AI model, which may mistake the artifact for a lesion or ignore a useful region.
That is why analysis pipelines start with automatic quality control: they segment (outline pixel by pixel) the usable tissue and the defective areas. Several tools exist, but according to the authors few can actually be used by third parties, and comparisons across several datasets are rare. This is the gap the preprint targets. It was posted on arXiv (category eess.IV) and submitted to Medical Image Analysis, so it has not yet been peer reviewed.
The method
The authors test six public algorithms on two tasks: isolating usable tissue, then classifying artifacts (background, dark spots or foreign objects, edges and air bubbles, folds, out-of-focus areas, pen marks). The six tools are: GrandQC (a U-Net++ network with an EfficientNet-B0 encoder, that is, a U-shaped image-segmentation network with a lightweight feature extractor), HistoKit (threshold-based tissue detection plus GrandQC's artifact module, run with overlapping windows), HistoQC (image metrics and classical classifiers), HistoROI (a ResNet18 classifier on small tiles, trained with a human in the loop), the work of Patil and colleagues (U-Net) and that of Schreiber and colleagues (plain thresholding in a colour space, using Otsu's method).
Three datasets serve as reference. Dataset A has 13 annotated regions from TCGA and CAMELYON16 (two public collections), with binary tissue-versus-rest segmentation. Dataset B has 567 annotated regions (the text mentions 574 images after filtering) from one Austrian centre, rebuilt from patch-level annotations. Dataset C, new, has 23 regions from 20 slides, drawn from TCGA (six organs) and the Compass NMD project (muscle), annotated by a doctoral student in ImageJ; it will be released on Zenodo upon publication.
Metrics are computed per pixel: the Dice index (the overlap between predicted and reference areas, from 0 to 1, where 1 is a perfect match), sensitivity (the share of true tissue pixels recovered) and the false discovery rate, or FDR (the share of pixels predicted as tissue that are not). Statistics use the Friedman test (which compares several methods on the same images) followed by pairwise Wilcoxon tests with a Benjamini-Hochberg correction, which limits the risk of false conclusions when many comparisons are made.
The results
ML metrics. Differences between methods are statistically significant. HistoKit and GrandQC are strong on all three datasets for tissue detection. HistoKit is significantly better than GrandQC on dataset B (Dice and FDR) and dataset C (all metrics), partly because GrandQC processes non-overlapping tiles that leave visible boundary artifacts. For artifact segmentation, GrandQC and HistoKit clearly outperform Patil and colleagues, which often confuses artifacts with blur or labels folds as tissue; HistoKit mainly improves folds, edges and out-of-focus areas on some datasets. The hardest classes for every method are dark spots and out-of-focus regions. HistoQC does worst: it often misses tissue, which mechanically lowers its FDR without making it better.
Computing time, measured on a single 0.98 GB image, changes the reading. Per slide, HistoROI takes 0.14 minutes, Schreiber 0.20, Patil 0.25, HistoQC 0.32, GrandQC 0.78 and HistoKit 6.18 (at 75% overlap). Overlap makes the difference: for HistoKit's artifact step, 0% overlap costs 0.41 minutes, 75% costs 2.02 and 90% costs 10.71. The authors suggest an overlap somewhat below 90%. A post-processing step (morphological opening with a 5-pixel disk, hole filling, an area threshold of 0.025% of the slide) significantly improves Dice and sensitivity on datasets A and B, with no effect on C.
Clinical translation. There is none, and that is the point to retain: the paper reports no diagnostic sensitivity, no patient-level false positives and no effect on a pathologist's reading. A better Dice on folds does not say how many slides would be wrongly rejected, how many artifacts would reach downstream diagnostic models, or what rescanning would cost. These conversions cannot be computed from the preprint's data.
What is good
A real comparison, with excluded tools named. The authors list the tools they could not test and why: PathProfiler fails to run, SliDL has inaccessible weights, SlideInspect, Smit and colleagues, GuardQC, AIRAQc and DiffusionQC have no usable code or weights, and GrandQC+ has no public scripts. This inventory documents the real state of reproducibility in the field.
Appropriate statistics. Friedman, then corrected Wilcoxon, with effect sizes: this is not the norm in the subfield, where averages are often compared without testing.
A measure of computing cost. Reporting 0.41 to 10.71 minutes depending on overlap shows a real trade-off between quality and time that method papers often omit. The analysis code, HistoKit and GrandQC are on GitHub, and dataset C will be released on Zenodo.
What is less good
Tiny annotated sets and fragile ground truth. Dataset A has 13 regions, dataset C 23 regions from 20 slides annotated by one trainee: potential label noise, with no inter-annotator agreement reported in the part read. Dataset B has inconsistent background labels, which the authors work around by excluding them from the reference. Dataset A does not annotate blur. Non-significant differences on C for some artifacts, such as edges, may simply reflect the small number of images.
Availability-driven comparator bias. Seven tools were dropped for lack of code or weights, HistoQC could not run on dataset B or read the magnification of three images in dataset A, and HistoROI did not run on B. The ranking therefore covers tools that work with little effort, not the full state of the art. One independence point also needs checking: the HistoKit repository sits under the account ZAEDPolSl, whose name suggests the same Silesian group as the authors. The page read does not say whether HistoKit is their own tool; if it is, a benchmark crowning the authors' own tool calls for confirmation by a third team. The authors declare no competing interests.
A pixel metric far from real use. Pixel-level Dice measures the quality of the map, not its usefulness. The authors acknowledge they do not evaluate the effect on downstream tasks (cell segmentation, tumour detection), that they tested only H&E, and that freezing artifacts are not covered. Population bias also applies to scanners and laboratories: GrandQC relies on data from one Austrian centre, and a quality-control tool trained this way may degrade at other centres, which only a large multicentre test would show. Finally, one prediction error can propagate to many small tiles at high magnification.
What it changes
For the research community. The paper provides a reproducible comparison base and a practical rule: tile overlap and morphological post-processing matter as much as architecture. Once released, the public dataset C will add a muscle test bed, which is rare in this field. It will need to be enlarged and annotated by several pathologists.
For clinicians and laboratories. For a laboratory digitising its slides, the message is modest: GrandQC or HistoKit look like good starting points, with a computing cost to weigh, and no tool replaces validation on one's own slides and scanners. There is no effect on diagnostic practice in the short term.
For patients and the public. The effect is indirect: better quality control reduces the risk that a diagnostic model is fooled by a slide defect, but this paper does not measure that risk.
Going further
Preprint: Merta J., Mendelsohn D., Schoser B., Polanska J., Marczyk M., "Benchmarking and Improvement of Artifact Segmentation Methods in H&E-Stained Whole Slide Imaging", arXiv:2610.24037, submitted to Medical Image Analysis, licence CC BY-NC-SA 4.0 (arxiv.org/abs/2610.24037). Affiliations: Silesian University of Technology (Gliwice, Poland), Friedrich-Baur-Institute, LMU Klinikum (Munich), Yale Cancer Center. Funding: CoMPaSS-NMD project, Horizon Europe, grant agreement 101080874. Code: benchmark analyses, HistoKit, GrandQC.
Editorial transparency note: this article was produced by Tatakoto's automated editorial pipeline from the HTML version of the arXiv preprint; the last 5,000 characters of the text could not be read, and figures are those reported in the part read. The English, Spanish and Chinese versions are AI-generated. Found an error? Contact us.