DINOCell: Self-supervised Pretraining of Cell Segmentation Models

DINOCell is an automated cell segmentation model for live cell microscopy images. Through initialization with DINOv2 weights, pretrained on 1.2B images, and domain-adaptation on 130k unlabeled cell images, DINOCell achieves unparalleled performance across a wide variety of cell types and microscope conditions.

Key Features

• State-of-the-art Performance: Outperforms existing methods like Cellpose-SAM and SAMCell on both test-set and zero-shot cross-dataset evaluation
• Zero-shot Generalization: Works on novel cell types and microscopes not seen during training
• Vision Transformer Architecture: Leverages DINOv2's ViT-based encoder pretrained on 1.2B images for robust image representations, which are further tuned for microscopy through domain-adaptation on 130k unlabeled cell images.
• Flows Regression: Predicts Cellpose-style Flows instead of binary masks, enabling better separation of densely packed cells

Performance

DINOCell demonstrates superior performance in both test-set and zero-shot cross-dataset evaluation:

LIVECell Test-Set Performance

Zero-Shot Cross-Dataset Performance

Quick Start

Installation

# Install from PyPI (recommended)
pip install dinocell

# Or install from source
git clone https://github.com/kadenstillwagon/DINOCell.git
cd DINOCell
pip install -e .

Basic Usage

import matplotlib.pyplot as plt
from dinocell import segment

#Set Image Path
img_path = 'demo_images/LiveCell_test_image.png'

#Get Model Output
output_segmentations = segment(img_path)

#Visualize
plt.imshow(output)
plt.show()

Command Line Interface

# Segmentation
dinocell segment image.png --output results/

Citation

If you use DINOCell in your research, please cite our paper:

Stillwagon K, VandeLoo AD*, Magondu B, Forest C.R. (2026) Self-supervised Pretraining of Cell Segmentation Models. https://arxiv.org/abs/2604.10609.

  @misc{stillwagon2026selfsupervisedpretrainingcellsegmentation,
        title={Self-supervised Pretraining of Cell Segmentation Models}, 
        author={Kaden Stillwagon and Alexandra Dunnum VandeLoo and Benjamin Magondu and Craig R. Forest},
        year={2026},
        eprint={2604.10609},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2604.10609}, 
  }

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Email: kstillwagon26@gatech.edu

License

This project is licensed under the MIT License - see the LICENSE file for details.

Institutions

This work was developed at:

• Georgia Institute of Technology
  • School of Biological Sciences
  • School of Computer Science
  • Department of Biomedical Engineering
  • School of Mechanical Engineering

Acknowledgments

• Meta AI for the original DINOv2
• The open-source community for tools and datasets
• Georgia Tech for computational resources
• All contributors and users of DINOCell