RedDino: A Foundation Model for Red Blood Cell Analysis

RedDino is a self-supervised Vision Transformer foundation model specifically designed for red blood cell (RBC) image analysis, as presented in the paper RedDino: A foundation model for red blood cell analysis.

It leverages a tailored version of the DINOv2 framework, trained on a meticulously curated dataset of 1.25 million RBC images from diverse acquisition modalities and sources. This model excels at extracting robust, general-purpose features for downstream hematology tasks such as shape classification, morphological subtype recognition, and batch-effect–robust analysis.

Unlike general-purpose models pretrained on natural images, RedDino incorporates hematology-specific augmentations, architectural tweaks, and RBC-tailored data preprocessing, enabling state-of-the-art performance on multiple RBC benchmarks.

🧠 Developed by Luca Zedda, Andrea Loddo, Cecilia Di Ruberto, and Carsten Marr
🏥 University of Cagliari & Helmholtz Munich
📄 Preprint: arXiv:2508.08180
💻 Code: https://github.com/Snarci/RedDino

Model Details

Architecture: ViT-large, patch size 14
SSL framework: DINOv2 (customized for RBC morphology)
Pretraining dataset: Curated RBC images from 18 datasets (multiple modalities and sources)
Embedding size: 1024
Intended use: RBC morphology classification, feature extraction, batch-effect–robust analysis Notes:
RBC-specific training strategy including removal of KoLeo regularizer and Sinkhorn-Knopp centering.
Training on smear patches (not only single cells) to enhance cross-source generalization.

Example Usage

from PIL import Image
from torchvision import transforms
import timm
import torch
# Load model from Hugging Face Hub
model = timm.create_model("hf_hub:Snarcy/RedDino-large", pretrained=True)
model.eval()
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
# Load and preprocess image
image = Image.open("path/to/rbc_image.jpg").convert("RGB")
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225]),
])
input_tensor = transform(image).unsqueeze(0).to(device)
# Extract features
with torch.no_grad():
    embedding = model(input_tensor)

Model Variants

RedDino comes in three sizes to suit different computational requirements and performance needs:

Model Variant	Embedding Size	Parameters	Usage
RedDino-small	384	22M	`timm.create_model("hf_hub:Snarcy/RedDino-small", pretrained=True)`
RedDino-base	768	86M	`timm.create_model("hf_hub:Snarcy/RedDino-base", pretrained=True)`
RedDino-large	1024	304M	`timm.create_model("hf_hub:Snarcy/RedDino-large", pretrained=True)`

Choose the variant that best fits your computational budget and performance requirements. Larger models generally provide richer feature representations at the cost of increased computational overhead.

Benchmark Results

RedDino was benchmarked on major RBC classification datasets—including Elsafty, Chula, and DSE—outperforming state-of-the-art baselines such as ResNet50, DinoBloom, and DINOv2.

Model	Dataset	Metric	Linear Probing (wF1)	1-NN (wF1)	20-NN (wF1)
ResNet50	Elsafty	Weighted F1	77.6 ± 8.1	64.3 ± 4.8	66.2 ± 4.9
DinoBloom-S	Elsafty	Weighted F1	83.2 ± 8.2	73.1 ± 5.1	76.5 ± 4.2
DINOv2 (small)	Elsafty	Weighted F1	82.1 ± 8.2	73.5 ± 4.8	77.2 ± 4.6
RedDino small	Elsafty	Weighted F1	86.0 ± 7.0	76.8 ± 4.9	80.0 ± 4.5
RedDino base	Elsafty	Weighted F1	88.1 ± 4.9	78.8 ± 3.6	82.6 ± 2.8
RedDino large	Elsafty	Weighted F1	88.5 ± 5.5	78.5 ± 4.6	81.6 ± 4.7

On Chula and DSE datasets, RedDino consistently surpassed all other models in feature quality (linear probing) with average improvements of 2–4% over prior approaches in key metrics.

Highlights

Foundation model for RBC analysis trained on the largest available multi-source RBC image set: 1.25M+ images, using advanced CellPose-based instance segmentation and patch extraction.
DINOv2-based self-supervised learning for label-efficient pretraining and robust, transferable features.
Model architecture and key innovations:
- Patch-based training (224×224 px) shown to outperform single-cell training.
- Novel data augmentation via Albumentations (32 pixel-level strategies).
- Removal of the Koleo regularizer and adoption of Sinkhorn-Knopp centering for improved representation in RBC-specific domains.
- Suite of models (small, base, large) covering 22M–304M parameters.
Generalization: Strong adaptation across varied protocols, microscopes, and imaging sites. Demonstrated resistance to batch effects and out-of-domain variance.
Interpretability tools: PCA/UMAP visualizations reveal clustering by phenotype and batch, distinguishing abnormal cells (e.g., malaria, echinocytes).
Easy deployment: Models and code are available on GitHub and Hugging Face.

📝 Citation

If you use this model, please cite the following paper:

RedDino: A foundation model for red blood cell analysis
Luca Zedda, Andrea Loddo, Cecilia Di Ruberto, Carsten Marr — 2025
Preprint: arXiv:2508.08180. https://arxiv.org/abs/2508.08180

@misc{zedda2025reddinofoundationmodelred,
      title={RedDino: A foundation model for red blood cell analysis}, 
      author={Luca Zedda and Andrea Loddo and Cecilia Di Ruberto and Carsten Marr},
      year={2025},
      eprint={2508.08180},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.08180}, 
}

Summary

RedDino is the first family of foundation models tailored for comprehensive red blood cell image analysis, using large-scale self-supervised learning to set new performance benchmarks and generalization standards for computational hematology. Models and pretrained weights are available for research and practical deployment.

Snarcy
/

RedDino-large