DINOv3
Paper • 2508.10104 • Published • 319
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is approved automatically. By requesting access, you share your Hugging Face username and email with HistoForge. We may get in touch to ask for your feedback.
Log in or Sign Up to review the conditions and access this model content.
Lizard Colon Nuclei (H&E)
Segments individual nuclei in human colon hematoxylin-and-eosin (H&E) images and assigns one of six cell-type classes to each nucleus. Returns nucleus objects, per-class counts and a total classified-nucleus count.
Experimental research model; not clinically validated.
Using the module
In HistoForge, open Modules, choose Install, and enter
histoforge/lizard-colon-nuclei. Then select the installed module for your
analysis regions.
Use brightfield RGB colon images with calibrated pixel size. The working resolution is 0.5 micrometres per pixel, approximately the scale reported for Lizard. HistoForge resamples calibrated inputs to this resolution. Review the contours and class assignments before interpreting or exporting counts, especially for rare granulocytes and densely packed epithelium.
| Output class | What it marks |
|---|---|
| Neutrophil | A nucleus classified as neutrophil. |
| Epithelial | A nucleus classified as epithelial. |
| Lymphocyte | A nucleus classified as lymphocyte. |
| Plasma | A nucleus classified as plasma cell. |
| Eosinophil | A nucleus classified as eosinophil. |
| Connective | A nucleus assigned the release's connective-tissue class. |
Background is not returned as an object. The module does not separate malignant from benign epithelial nuclei or identify finer immune subtypes.
Measurements
Counts use the current classified objects in each selected region, including manual additions, deletions and class corrections.
| Measurement | Meaning | Unit |
|---|---|---|
| Neutrophil Nuclei | Number of Neutrophil objects. | nuclei |
| Epithelial Nuclei | Number of Epithelial objects. | nuclei |
| Lymphocyte Nuclei | Number of Lymphocyte objects. | nuclei |
| Plasma Nuclei | Number of Plasma objects. | nuclei |
| Eosinophil Nuclei | Number of Eosinophil objects. | nuclei |
| Connective Nuclei | Number of Connective objects. | nuclei |
| Total Classified Nuclei | Sum of the six counts above. | nuclei |
An empty valid region returns zero. Failed regions are not zero-count observations. Counts sum across regions; use non-overlapping regions to avoid double counting. Other classes and unclassified objects are excluded. See Measurement definitions for exact class matching.
Examples
The three fields were selected by deterministic pseudorandom sampling, one each from CRAG, DigestPath and GlaS, before inspecting images or scores. They were used for checkpoint-selection validation, but not for fitting model parameters. They are not independent test examples. Each figure shows original | released annotation | prediction over identical pixels.
Performance
Measured on all 70 Lizard fold-2 validation images, containing 154,259 annotated nuclei across CRAG, DigestPath and GlaS. This population selected the checkpoint and initialization, so these are development results, not an independent accuracy estimate. No test-set result is claimed.
The table measures the packaged weights using the current finite-image HistoForge backend on GH200, with BF16 inference and emitted classed polygons. The 95% intervals use 2,000 bootstrap samples of source groups.
| Metric | Value | 95% interval |
|---|---|---|
| Six-class panoptic quality (mPQ+) | 0.512 | 0.491 to 0.531 |
| Binary panoptic quality | 0.688 | 0.675 to 0.700 |
| Foreground Dice | 0.829 | 0.822 to 0.836 |
| Aggregated Jaccard index (AJI) | 0.680 | 0.671 to 0.690 |
| Segmentation score at IoU 0.50 (AP50) | 0.735 | 0.721 to 0.748 |
| Mean segmentation accuracy (mSA) | 0.429 | 0.415 to 0.442 |
| End-to-end object micro-F1 | 0.780 | 0.770 to 0.791 |
| End-to-end object macro-F1 | 0.634 | 0.609 to 0.657 |
mPQ+ averages the six class-specific panoptic qualities. Binary PQ ignores class. Object F1 counts missed, extra and incorrectly classified nuclei; a correct object requires intersection-over-union (IoU) above 0.5 and the correct class. It is not classification accuracy restricted to matched nuclei. AP50 is TP/(TP+FP+FN) at IoU 0.50, not a confidence-ranked curve area; mSA averages this quantity across IoU thresholds 0.50 to 0.95. AJI, AP50 and mSA are image means, while PQ, object F1 and foreground Dice pool their counts.
| Output class | Annotated nuclei | Predicted nuclei | Object F1 | Class PQ+ |
|---|---|---|---|---|
| Neutrophil | 1,839 | 1,168 | 0.384 | 0.290 |
| Epithelial | 74,941 | 67,501 | 0.817 | 0.635 |
| Lymphocyte | 34,732 | 36,777 | 0.816 | 0.696 |
| Plasma | 8,218 | 6,964 | 0.604 | 0.520 |
| Eosinophil | 1,138 | 852 | 0.452 | 0.335 |
| Connective | 33,391 | 29,924 | 0.729 | 0.594 |
Rare-cell uncertainty is substantial: neutrophil PQ+ has a 95% interval of 0.208 to 0.364, and eosinophil PQ+ 0.285 to 0.392. These intervals do not account for checkpoint-selection optimism or variation across training seeds.
For context, the Lizard authors reported the following three-fold cross-validation results in Table 4. Their population, fold rotation and multiclass-PQ aggregation differ from this single-split evaluation. These quoted values are not an author-model rerun or a same-set ranking.
| Paper model | Binary Dice | Binary PQ | Multiclass PQ |
|---|---|---|---|
| U-Net | 0.735 ± 0.028 | 0.515 ± 0.033 | 0.265 ± 0.013 |
| Micro-Net | 0.786 ± 0.004 | 0.522 ± 0.015 | 0.264 ± 0.012 |
| HoVer-Net | 0.828 ± 0.008 | 0.624 ± 0.013 | 0.396 ± 0.022 |
The historical research evaluator measured this same checkpoint at mPQ+ 0.512 and binary PQ 0.688. Its precision, voting and decoding implementation differ. Input normalization was numerically identical on these 70 images. The table above reports current serving behavior; it does not establish live L40S or installed QuPath equivalence.
Limitations
- The evaluation population is development validation used for checkpoint selection. Independent performance on the reserved fold, other institutions, scanners, staining protocols and organs is not established.
- Neutrophils and eosinophils have relatively few annotated examples. Treat their counts and uncertainty separately from abundant epithelial nuclei.
- Lymphocyte versus plasma-cell labels can be difficult even for annotators. The released eosinophil contours can include visible cytoplasm.
- Missed, extra, merged, split or misclassified nuclei can bias every count. The model provides no calibrated confidence or abstention score.
- Finite-image evaluation does not establish whole-slide stitching, live L40S behavior or end-to-end QuPath persistence and export.
- The module does not measure cell density, tissue area, a clinical score, prognosis or treatment response.
Model and training
One network predicts Cellpose flows, foreground probability and seven class logits: background plus the six reported cell types. Its backbone is DINOv3 ViT-L/16, with a multiscale spatial adapter and native-resolution readout. The trunk starts from the official DINOv3 weights. The adapter and segmentation readout start from nucleus pretraining on TNBC, NuInsSeg and LyNSeC; the classification head starts fresh. All network parameters were fine-tuned using 512-pixel crops at seed 42. Validation selected step 1,500, corresponding to 24,000 sampled crop exposures.
Training used 70 fold-1 images containing 149,371 annotated nuclei. Validation used 70 fold-2 images with 154,259 annotated nuclei. The reserved 98-image fold was not evaluated. No all-data refit was performed.
| Component | Purpose | Artifact |
|---|---|---|
| Nucleus model | RGB image to classed nucleus instances | Weights |
| Module recipe | Image scale, tiling and output classes | recipe.json |
| Measurement script | Classified nuclei to seven counts | metrics.groovy |
Historical training and current serving use separately pinned implementations. See the reproduction guide for exact sources, inputs, commands and limits. Source access requires repository permission; public clean-room reproduction is not yet available.
Licences and citations
- Weights: DINOv3 License. This model uses DINOv3.
- Measurement script and configuration: Apache-2.0.
- Example images and released annotations: CC BY-NC-SA 4.0, for research and non-commercial use; see Example credits.
- Training data: Training datasets and licences. Training datasets are not included in the module.
Dataset: Graham et al. (2021), Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification. Paper.
Foundation model: Siméoni et al. (2025), DINOv3. arXiv:2508.10104.
Instance-decoding method: Stringer et al. (2021), Cellpose: a generalist algorithm for cellular segmentation. doi:10.1038/s41592-020-01018-x.
Model tree for histoforge/lizard-colon-nuclei
Base model
timm/vit_large_patch16_dinov3.lvd1689m