Image credits
Lizard: cells
The cell image is from the Lizard dataset, version 1, released by Simon Graham and colleagues under CC BY-NC-SA 4.0 for research and non-commercial use. The cropped image with HistoForge’s predicted nucleus outlines and class colours retains the same licence.
The source is H&E-stained colon tissue, image dpath_36 from DigestPath.
The displayed 885 × 885 crop starts at pixel coordinates x = 84, y = 0.
The source tissue is overlaid in QuPath with predicted nucleus boundaries and
translucent class colours. The crop is not resized, stretched or padded.
Predicted class |
Overlay colour |
|---|---|
Epithelial |
Yellow-orange ( |
Lymphocyte |
Pink ( |
Plasma |
Light blue ( |
Connective tissue |
Green ( |
Eosinophil |
Vermilion ( |
These colours mark nuclei, not whole cells. No nuclei are predicted as neutrophils in this crop. Some nuclei are missed or misclassified, including plasma cells.
This validation crop was deliberately selected for illustration after comparing predictions with the released annotations. It was held out from fitting model parameters, but validation scores contributed to model selection. This is an experimental model illustration, not independent test evidence or a claim of diagnostic accuracy. The dataset authors did not provide or endorse these predictions.
Related paper: Graham et al., Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification (2021).
Da-Mi Kim: tissue structures
The structure image shows PAS-stained rat testis from Da-Mi Kim’s dataset, licensed under CC BY 4.0. The complete 2048 × 2048 field is retained. HistoForge adds outlines and translucent stage colours to 11 detected tubule cross-sections.
The source field is 25_10_jpg.rf.2975e1a994d910c283fcf4edd854284c.jpg.
It belongs to development validation material used in model selection, not an
independent test. The predictions come from HistoForge’s experimental rat
seminiferous-tubule staging model, not the dataset authors’ model. The authors
did not provide or endorse these predictions.
The image has no physical scale bar: its demonstration pixel size is an estimate, not verified slide calibration.
The Metrics view is a screenshot of HistoForge’s Project tab and actual measurement table. It shows tubule area, minor and major axes, and circularity. The screenshot is retained at its original 1714 × 1724 resolution. The physical scale in this demonstration has not been verified, so the displayed area and length values are illustrative, not calibrated scientific measurements.
Related paper: Kim, Rho, Wee and Son, Automated spermatogenic staging in periodic acid-Schiff-stained testes of Sprague–Dawley rats using a deep learning model for normal and atrophied tissues (2026).
Da-Mi Kim: within structures
This image shows PAS-stained rat testis from Da-Mi Kim’s dataset, under
CC BY 4.0.
The source field is 30_14_jpg.rf.4e63e8ab51c4cd6a6706e81ce6ebcac0.jpg.
The 1536 × 1536 crop starts at x = 128, y = 256 and is not resized or padded.
It comes from development validation material, not an independent test.
Overlay |
Colour |
|---|---|
Detected tubule outlines |
Green ( |
Predicted nuclei |
Red ( |
Predicted lumens |
Amber ( |
Only overlays inside the tubules are shown, without cell classifications. The cell model ran separately on a crop around each detected tubule, including those at the image edge. The lumen model ran on the full field shown here. Both overlays are clipped to tubule outlines for display. The image was exported from QuPath with filled nucleus masks and outlines. For this illustration, lumen overlays are hidden in the central tubule and the tubule cut off at the top-center. The main lumen overlay in the top-right corner tubule is retained, but a small fragment beside it is hidden. The saved model predictions are unchanged.
These are unreleased research models trained on H&E tissue. This PAS example illustrates the workflow, not independently measured accuracy. It does not imply that the models are available to download. There is no physical scale bar because the demonstration pixel size has not been verified.
RINGS: tissue regions
The landing page shows prostate tissue from the RINGS algorithm dataset, version 1, by Massimo Salvi, Martino Bosco, Luca Molinaro, Alessandro Gambella, Mauro Giulio Papotti, Udyavara Rajendra Acharya and Filippo Molinari (2021), licensed under CC BY 4.0.
The colours show model predictions: blue for gland-only regions, pink for tumour-only regions, and orange where gland and tumour predictions overlap. Uncoloured tissue retains its original appearance. These are region masks, not outlines of individual glands.
Image P4_C10_16_9_7 comes from the dataset’s training material and was assigned to our
development validation split. It illustrates model output, not independent performance
or a clinical assessment. The dataset authors did not provide or endorse these predictions.
Changes to the source image: resized and overlaid with model predictions. The full tissue field is retained. The landing image was extracted from the prediction panel of the original–annotation–prediction comparison without changing its colours or masks.
Related paper: Salvi et al., A hybrid deep learning approach for gland segmentation in prostate histopathological images (2021).
Whole slide inference
The slide is CMU-1.svs from the
OpenSlide Aperio collection,
released under CC0 1.0,
as listed by QuPath.
Four tissue regions were drawn by hand in QuPath. One is highlighted in yellow; this is selection, not model output. HistoForge’s Annotations by class scope includes all four Tissue region annotations, regardless of which is highlighted. This view shows the controls before a module is assigned or analysis is run.
The two actual windows were cropped from one screenshot and arranged side by side. The desktop background was removed; window pixels and proportions are unchanged.
Module library
The Modules-tab image shows installed modules and documentation for the selected DeepLIIF Ki-67 module. It is an interface example, not a model-performance result or evidence of a completed publication workflow.
Training interface
The Training view shows DeepLIIF selected and the controls for choosing training and validation data. No training regions or data split have been prepared in this capture; it does not show a training run or results.