A tissue and cell-level annotated H&E and PD-L1 histopathology image dataset in non-small cell lung cancer
Paper • 2507.16855 • Published
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.
IGNITE NSCLC Tissue Segmentation (H&E)
Map tissue compartments in H&E-stained non-small cell lung cancer (NSCLC). The module returns coloured tissue regions in QuPath and their areas in mm². It segments tissue, not individual cells or nuclei.
Using the module
Module reference: histoforge/ignite-nsclc-he-tissue.
The workflow uses a calibrated H&E image and selected analysis regions in QuPath. HistoForge samples the image at 0.5 µm/pixel and uses its standard tiled inference. Review the labelled regions, correct them if needed, then calculate or export measurements. Area measurements reflect the current objects, including your edits.
| Output class | What it marks |
|---|---|
| Tumor epithelium | Tumour epithelial tissue. |
| Reactive epithelium | Epithelium showing reactive changes. |
| Stroma | Connective tissue. |
| Inflammation | Inflammatory-cell-rich regions. |
| Alveolar tissue | Alveolar lung tissue. |
| Fatty tissue | Adipose tissue. |
| Necrotic tissue | Necrotic (dead) tissue. |
| Erythrocytes | Red-blood-cell regions. |
| Bronchial epithelium | Epithelium lining bronchi. |
| Mucus/Plasma/Fluids | Mucus, plasma and other fluids, grouped as one class. |
| Cartilage/Bone | Cartilage and bone, grouped as one class. |
| Macrophages | Macrophage regions. |
| Muscle | Muscle tissue. |
| Liver | Liver tissue. |
| Keratinization | Keratinized tissue. |
Background is predicted internally and is not returned as a tissue object. Connected regions are display geometry, not cell counts.
Measurements
Measurements are calculated within each selected region of interest (ROI).
| Measurement | Meaning | Unit |
|---|---|---|
| IGNITE Analyzed Area | Area of the analysis ROI, including background. | mm² |
| IGNITE Classified Tissue Area | Area covered by any of the 15 tissue classes. | mm² |
| Tumor Epithelium Area | Area marked as Tumor epithelium. | mm² |
| Reactive Epithelium Area | Area marked as Reactive epithelium. | mm² |
| Stroma Area | Area marked as Stroma. | mm² |
| Inflammation Area | Area marked as Inflammation. | mm² |
| Alveolar Tissue Area | Area marked as Alveolar tissue. | mm² |
| Fatty Tissue Area | Area marked as Fatty tissue. | mm² |
| Necrotic Tissue Area | Area marked as Necrotic tissue. | mm² |
| Erythrocytes Area | Area marked as Erythrocytes. | mm² |
| Bronchial Epithelium Area | Area marked as Bronchial epithelium. | mm² |
| Mucus/Plasma/Fluids Area | Area marked as Mucus/Plasma/Fluids. | mm² |
| Cartilage/Bone Area | Area marked as Cartilage/Bone. | mm² |
| Macrophages Area | Area marked as Macrophages. | mm² |
| Muscle Area | Area marked as Muscle. | mm² |
| Liver Area | Area marked as Liver. | mm² |
| Keratinization Area | Area marked as Keratinization. | mm² |
Tissue areas are clipped to the analysis ROI; same-class overlaps are counted once. An absent class has an area of zero. Areas are summed across non-overlapping analysis ROIs and images. See measurement definitions for calibration, geometry and aggregation details.
Examples
Three TEST crops from different patients, selected by deterministic pseudorandom sampling from fully annotated tissue crops; none of these patients was used for training. Each shows original image | annotation overlay | model prediction from the reported research evaluation, with a shared colour legend.
Images and annotations: Spronck et al., IGNITE v1, CC BY-NC-SA 4.0. Tissue colours and the prediction overlay have been added. Example selection and credits.
Performance
Both scoring protocols below use the same saved predictions for each model on 139 official evaluation regions from 34 patients. Dice/F1 measures pixel agreement, not cell detection or classification.
HistoForge evaluation convention
| Model | Macro tissue Dice | Micro tissue Dice |
|---|---|---|
| This model | 0.7225 | 0.7931 |
| Released IGNITE five-model nnU-Net ensemble, rerun | 0.6733 | 0.7870 |
Only pixels marked unannotated in the ground truth are ignored. Tissue predictions on annotated background and missed tissue count as errors. Confusion counts are pooled across all regions; macro Dice then averages the 15 tissue-class scores equally, while micro Dice pools their true-positive, false-positive and false-negative counts.
Per-class pixel F1 (Dice)
The same scoring convention gives the following tissue-class breakdown.
| Tissue class | This model | Released IGNITE ensemble, rerun |
|---|---|---|
| Tumor epithelium | 0.8950 | 0.8786 |
| Reactive epithelium | 0.5188 | 0.5621 |
| Stroma | 0.7734 | 0.7800 |
| Inflammation | 0.6748 | 0.6831 |
| Alveolar tissue | 0.5975 | 0.5655 |
| Fatty tissue | 0.8920 | 0.8313 |
| Necrotic tissue | 0.7744 | 0.7761 |
| Erythrocytes | 0.6985 | 0.7489 |
| Bronchial epithelium | 0.7705 | 0.7316 |
| Mucus/Plasma/Fluids | 0.4542 | 0.4521 |
| Cartilage/Bone | 0.6412 | 0.4516 |
| Macrophages | 0.7789 | 0.7715 |
| Muscle | 0.6245 | 0.3912 |
| Liver | 0.9537 | 0.9608 |
| Keratinization | 0.7905 | 0.5154 |
Authors' evaluation protocol
| Model / evidence | All-tissue micro Dice/F1 | Grouped-tissue micro Dice/F1 |
|---|---|---|
| This model | 0.7943 | 0.8222 |
| Released IGNITE five-model nnU-Net ensemble, rerun | 0.7870 | 0.8134 |
| IGNITE paper (2025), reported values | 0.79 | 0.81 |
The authors' evaluator additionally excludes pixels predicted as unannotated and pixels annotated as background. Its grouped score combines the tissue classes into tumour, stroma/inflammation, necrosis, macrophages and remaining tissue. This is a segmentation score, not a clinical tumour-infiltrating lymphocyte measurement. The paper row is quoted; the ensemble row is our rerun of the released weights.
The evaluation set was reused during the study, so these are descriptive comparisons, not independent confirmation of performance. Superiority over the ensemble was not established. These scores use the research predictions, not measured HistoForge-runtime performance; HistoForge's different tile placement and blending have not been benchmarked.
Limitations
- Training and evaluation cover the IGNITE H&E NSCLC domain. Rare classes, ambiguous boundaries, staining differences and tissue outside that domain require review.
- Unannotated dataset pixels are unknown, not confirmed background.
- This module does not analyse PD-L1 staining or calculate clinical biomarker scores. It is a research tool, not a diagnostic system.
Models and training
The model uses a DINOv3 ViT-L/16 encoder with a multiscale adapter and a 16-channel semantic head. It was initialized from official DINOv3 weights and fine-tuned on IGNITE v1. The final model was trained on 269 H&E regions from 48 patients, for 2,813 optimizer steps.
Data preparation and reproduction guide.
Licences and citations
- IGNITE dataset v1 (Spronck et al., 2025): CC BY-NC-SA 4.0.
- Model weights: DINOv3 licence.
- Measurement script: Apache-2.0.