fidoptNuInsSeg, nuclei outlined by hand on H&E
A 512-pixel H&E patch of one whole slide of a human or mouse organ, every nucleus outlined by hand and the areas too ambiguous to outline marked apart (NuInsSeg, Medical University of Vienna).
Read before comparing (8)
- Slides. 31 whole slides, one an organ, scanned at 40x on a TissueFAXS: 23 human organs from the teaching collection of the Medical University of Vienna's Center for Anatomy and Cell Biology, and eight organs of 8-week-old male C57BL/6J mice. A senior cell biologist chose the fields of each slide (9 to 47 human, 6 to 42 mouse); each patch is the centre of one field.
- Subjects. The deposit names no donor: each human slide is taken as a subject, though two organs of the collection may come from one body. The mouse slides are one case because their animals are not told apart.
- Ignore. The provider's ambiguous areas, where the authors judged no outline deterministic, are taken out of the truth and the prediction before scoring and never shown to methods.
- Labels. The provider numbers each nucleus; prepare.py makes them one label, so touching nuclei become one region and the corpus scores no instances.
- Preparation. Ten human tissues were embedded in celloidin and cut at about 15 to 20 µm (jejunum, kidney, liver, oesophagus, tonsil, pancreas, placenta, salivary gland, spleen, tongue), the others in paraffin at 4 to 5 µm: thick sections show nuclei out of focus.
- Units. The files give no pixel size, so distances are in pixels and the surface tolerance is one pixel.
- Start. On the 10 test patches the nuclei cover 9.4% of a patch on average (0.9 to 31%, ambiguous areas out): the whole patch scores a Dice of 0.18, Otsu's dark pixels 0.42 and its bright ones 0.02, with a surface Dice of 0.10 at one pixel for the dark ones. TotalSegmentator has no histology task.
- Overlap. No other dataset of the corpus holds these slides; the record is also on Kaggle (ipateam/nuinsseg), which this copy does not use.
| Accuracy | Cost | |||||||
|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 px | Time s | RAM GB | GPU GB | Images |
| 1 | Otsu nucleiCPUapp | 0.636 ±0.184 | 0.310 ±0.147 | 56.87 ±51.16 | 5 | 0.8 | ||
The order averages each method's ranks on Dice, Surface Dice, HD95 px, so the first one is not always the best in every column. ± is the 95% interval of the mean over the 10 cases. Two neighbours are tied when a paired Wilcoxon test on the shared cases does not find the upper one better on Dice (5%): the places a method could hold are then shown under its own. The best value of each column is in colour. Click a row to see its images, a column title to sort by it. Time and memory measured on Laptop.
Under a licence that is not open
Methods whose code or weights are released for non-commercial use, or under share-alike or custom terms: shown for reference, ranked nowhere.
| Accuracy | Cost | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| # | Method | Dice | Surface Dice | HD95 px | Time s | RAM GB | GPU GB | Machine | Images |
| Cellpose (cyto3)CPUapptrained on CC-BY-NC-4.0 AND CC-BY-NC-SA-4.0 AND NOASSERTION dataolder run | 0.419 ±0.234 | 0.235 ±0.139 | 218.25 ±101.05 | 21 | 2.0 | Laptop | |||
| StarDist (H&E)CPUapptrained on CC-BY-NC-SA-4.0 AND CC-BY-4.0 dataolder run | 0.524 ±0.237 | 0.246 ±0.114 | 165.00 ±145.17 | 12 | 1.6 | Laptop | |||
Dice: how much the organ contours overlap, 1 is perfect. Surface Dice: the share of the contours' surfaces within a tolerance of each other, 1 is perfect. HD95: distance between the contours' surfaces.
Case by case
Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.
About the data
- Test cases
- 10 test slides drawn with seed 2026, then 10 development slides, one patch a slide. The pool is the 23 human slides, one an organ, and the eight mouse slides taken together as one case, since the deposit does not say which mice their organs came from.
- Truth
- contours by experts every nucleus of the patch outlined freehand in ImageJ by one of three students trained in cell biology, checked and corrected where needed by a senior cell biologist; the corpus merges the instances into one label
- Source
- zenodo.org/records/10518968
doi.org/10.5281/zenodo.10518968
doi.org/10.1038/s41597-024-03117-2
github.com/masih4/NuInsSeg - Licence
- CC BY 4.0 (the Zenodo record's licence field)
10 cases for tuning, never tested
human_muscle, human_tonsile, mouse, human_placenta, human_jejunum, human_pancreas, human_brain, human_tongue, human_testis, human_melanoma