Brain Region Segmentation — Mouse Brain
DINOv2-Large + UperNet model fine-tuned for semantic segmentation of mouse brain structures in Nissl-stained histological sections.
Model Details
| Attribute | Value |
|---|---|
| Architecture | DINOv2-Large (304M) + UperNet (38M) |
| Classes | 1,328 |
| Input Size | 518x518 |
| Training Data | Allen Brain Institute CCFv3 10um Nissl staining |
| mIoU (center-crop) | 74.8% |
| mIoU (sliding window) | 79.1% |
Usage
git clone https://github.com/Noel-Niko/histological-image-analysis
cd histological-image-analysis
make install
make download-models-mouse
make annotate-mouse IMAGES=/path/to/your/slides/
Paper
Transfer Learning for Ultra-Fine-Grained Brain Region Segmentation: An Ablation Study with DINOv2 + UperNet on 1,328 Allen Mouse Brain Atlas Structures
We conduct a systematic ablation study across 9 training runs on pixel-level segmentation of 1,328 brain structures from the Allen Mouse Brain Atlas CCFv3. The two most impactful interventions are backbone partial fine-tuning (+8.5% mIoU) and extended training from 100 to 200 epochs (+6.0% mIoU). Three attempted improvements — weighted Dice+CE loss, aggressive augmentation, and test-time augmentation — all degrade performance.
See paper.md in this repo for the full paper.
Citation
If you use this model, please cite the training data sources and the paper included in this repository.
Repository
Full source code, training notebooks, and all models: https://github.com/Noel-Niko/histological-image-analysis
Maintaining This Repo
To update model weights, papers, or this README:
cd histological-image-analysis
export HUGGING_FACE_TOKEN=hf_your_token_here
# Update model weights (Databricks or local):
jupyter notebook notebooks/upload_models_to_hf.ipynb
# Update papers + READMEs (local only):
jupyter notebook notebooks/upload_papers_to_hf.ipynb
- Downloads last month
- 19