About and citations

A set of counting aids for pathology, running entirely on the user's own machine.

For reference only. Not for diagnostic use. The pathologist reads the slide and is responsible for the conclusion.

Intended use

These tools help a pathologist count on microscope fields and whole-slide images: positive nuclei, membrane staining classes, and in situ hybridisation signal dots. They replace manual counting. They do not replace reading the slide.

They are used inside the Department of Pathology, University Medical Center Ho Chi Minh City, for research, teaching and quality improvement. Any figure must be interpreted together with the pathologist's own assessment of that same slide.

Disclaimer

Not intended for clinical, diagnostic or therapeutic purposes. This is not a medical device. It has no regulatory clearance, is not registered as an in vitro diagnostic device, and has not been clinically validated on an independent cohort. For research use only. Not for use in diagnostic procedures.

No substitute for professional judgement. Every figure here is for reference. The pathologist reads the slide again and is responsible for the final conclusion. No result from these tools may be the sole basis for a diagnosis, a category assignment or a treatment decision.

No warranty of accuracy or completeness. The software is provided as is. The authors accept no liability for any loss arising from its use, including when a result is wrong.

The thresholds were fitted on one laboratory's slides. A different scanner, stain, section thickness or focus shifts the numbers. Before using the tools on slides from another source, recalibrate and compare against hand counts.

The tools are still being adjusted. Thresholds and counting rules change at each calibration, so two versions can give two different numbers on the same slide. Record the date of the count alongside any result.

Do not enter patient-identifying information. File names, field names and exported label files can all travel further than intended; name them with a study code. Handling of slides and patient data follows hospital policy.

About us

An internal product of the Department of Pathology, University Medical Center Ho Chi Minh City (UMC). Built for the department's own counting work, not released commercially.

The counting rules and every threshold in these tools were fitted on the department's own slides together with hand counts by its pathologists: each nucleus outlined and each signal dot placed by hand, then the machine adjusted to match. Each calibration round, with the measurements before and after, is recorded in the project changelog.

The slide data used for calibration and as samples has had the label image, the barcode and the scanner details removed.

What the department and the developers undertake to do

A disclaimer is not a process. These commitments come with the tools:

The pathologist verifies before use. No machine figure enters a record or a report until the pathologist has checked it against that same slide. The figure reported is the pathologist's, not the machine's.
Internal validation for each marker before routine use. For every marker, the department compares the machine against hand counts on enough cases, records the difference and the agreement, and keeps that on file. A marker without such a file is used only for research and teaching.
Recalibration when conditions change. A change of scanner, stain, antibody or section thickness means recalibrating and comparing against hand counts again before further use.
Every change is logged. Each threshold change is recorded with its date, its reason and the measurement either side. Saved results carry the date of the count, because two versions can give two numbers.
Periodic review. A randomly chosen sample is re-counted by hand at intervals, to catch drift away from the calibrated level.
Reported errors are reproduced and measured before anything changes. Pathologists report a misread case through the feedback button or directly. Each report is reconstructed and measured first, never fixed on impression, and the outcome goes into the log.
Known limits are stated. Where the machine still reads wrongly is written openly in the changelog rather than hidden, with the measurements showing how far off it is.
No extension of scope without validation. Not used for markers, tissue types or slide sources that have not been calibrated.
Data stays on the machine. No upload step is added.

Terms of use

Free of charge. The tools are built on openly published models, libraries and literature, and are provided free in turn: free in, free out.
No redistribution. Do not copy, repost or share the tools, the source, the models or the sample data outside the department.
No derivative works. Do not modify them, do not re-tune the model, and do not use them as the basis for another product, commercial or otherwise.
Ask first. Any use beyond the Department of Pathology, University Medical Center Ho Chi Minh City requires the authors' written permission.
The user carries the risk. Results must be checked by a pathologist against the slide. Any error arising from not checking is the user's responsibility.

These terms cover the department's own work. Third-party components keep their own licences, some of which are broader: the bundled label file ignite-pdl1.json is under CC BY-NC-SA 4.0, so anyone may share and adapt that file non-commercially, with attribution and under the same licence. See the dataset and third-party sections below.

Privacy

Images and slides do not leave your machine. Reading the image, segmenting nuclei and counting signals all happen in the browser; there is no upload step. A working session's figures are stored in your own browser.

Datasets used to build the tools

Every external dataset used here is openly published and used within what its licence allows. The tools are not released commercially, so the non-commercial conditions below are respected.

IGNITE data toolkit (Radboudumc, Zenodo 17735903), CC BY-NC-SA 4.0. Used to train the tumour / non-tumour cell classifier (cell-type.onnx) and as the base for the label-learning layer. The label file bundled with the tools (ignite-pdl1.json, 8,000 balanced cells) is derived from IGNITE and therefore stays under CC BY-NC-SA 4.0, attributed inside the file itself and in the interface.
TIGER (Zenodo 6014422), CC BY-NC 4.0. Used only to evaluate the duct and DCIS detector. Not used for training and not bundled with the tools.
BCSS (Amgad et al. 2019), CC0, pixel-level regions on TCGA-BRCA H&E slides. Used only for evaluation and not bundled.
StarDist 2D_versatile_he model, BSD. Reused unmodified, not retrained and not fine-tuned.
Slides from the Department of Pathology, UMC. Used to calibrate thresholds and as sample slides. The label image, barcode and scanner details have been removed. Their use follows the hospital's internal rules on patient data.

No dataset whose terms forbid this kind of use was used, and no weights from any commercial product were used.

Third-party software

StarDist, nucleus segmentation model 2D_versatile_he, reused unmodified, BSD licence.
ONNX Runtime Web 1.20.1 (Microsoft), MIT licence, runs the models in the browser.
geotiff.js 3.0.5, MIT licence, reads .svs files and pyramidal TIFF.
QuPath used as a reference when checking nucleus segmentation. It is not part of the tools.

References

Diagnostic guidelines

Wolff AC et al. Human Epidermal Growth Factor Receptor 2 Testing in Breast Cancer: ASCO/CAP Clinical Practice Guideline Focused Update. J Clin Oncol 2018. Groups 1 to 5 for dual-probe in situ hybridisation; groups 2 to 4 require concurrent immunohistochemistry and a second observer. ascopubs.org
College of American Pathologists. HER2 breast update algorithms, 2023. Decision diagrams. documents.cap.org
Ventana INFORM HER2 Dual ISH DNA Probe Cocktail, Interpretation Guide (FDA P100027). Conventions for touching dots, clusters and doublets. accessdata.fda.gov
Nitta H et al. 2008. The brightfield dual-colour in situ hybridisation method, 98.9 % concordance with FISH. PMC2577627

Automated dual-ISH counting

CHERISH study, Virchows Archiv 2024 (HALO ISH, Indica Labs), 80 cases. Nucleus size filter 30 to 150 µm²; at least 469 cells for a ratio error under 0.1 and at least 953 cells for the copy number. PMC11950096
Gough A et al. Histopathology 2023. The algorithm raised inter-observer agreement from kappa 0.471 to 0.666. PubMed 37366040
Roche uPath HER2 Dual ISH. Counting inside a pathologist-annotated viable tumour region, confirmed by the pathologist. diagnostics.roche.com
SISH with StarDist and image processing, 2024. Nuclei need at least 2 CEP17 dots, at least 20 nuclei per image. PMC11623010
FISH with YOLOv3 and ResNet18, Scientific Reports 2023. Under-counts where signals cluster. PMC10275857
HER2-SISH segmentation with a Vision Transformer, 2024. PMC11593209
DISH with lightweight deep learning, medRxiv 2025. medRxiv

Nucleus segmentation and splitting

Schmidt U, Weigert M, Broaddus C, Myers G. Cell Detection with Star-Convex Polygons. MICCAI 2018. The StarDist model used here.
StarDist for H&E histopathology, IEEE 2022. ieeexplore.ieee.org
Graham S et al. HoVer-Net. Medical Image Analysis 2019. Separating touching nuclei with a centre and flow model. sciencedirect.com
Bankhead P et al. QuPath: Open source software for digital pathology image analysis. Scientific Reports 2017. Used as a reference.
Koyuncu CF et al. 2016, iterative h-minima. onlinelibrary.wiley.com
Splitting overlapping nuclei from outline concavity points. academia.edu
Iterative distance transform, EURASIP Journal on Image and Video Processing 2020. springeropen.com
Review of deep-learning nucleus segmentation, Medical Image Analysis 2024. sciencedirect.com

Invasive region

Invasive region segmentation on whole-slide images, 2024. PMC10778369

Citing these tools

Department of Pathology, University Medical Center Ho Chi Minh City. Pathology counting tools (internal software), 2026. tool-gpb-umc.pages.dev