DinoBloom: A Foundation Model for Generalizable Cell Embeddings in Hematology

DinoBloom was developed by Koch et al. and more information can be found on their GitHub repository and in the accompanying paper. This repository is fork of their HuggingFace repository. DinoBloom builds upon DINOv2 (Meta AI) and is trained on 13 diverse publicly available datasets of single cells from peripheral blood and bone marrow.

---

🧠 Model Variants

DinoBloom is available in four sizes:

Model	Feature Dim	Parameters	Checkpoint
DinoBloom-S	384	22M	pytorch_model_s.bin
DinoBloom-B	768	86M	pytorch_model_b.bin
DinoBloom-L	1024	304M	pytorch_model_l.bin
DinoBloom-G	1536	1136M	pytorch_model_g.bin

---

🔧 Requirements

pip install torch torchvision huggingface_hub

---

🚀 Usage

from huggingface_hub import hf_hub_download
import torch
import torch.nn as nn

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Choose variant: "s", "b", "l", or "g"
variant = "b"

# Configuration
variant_config = {
    "s": ("dinov2_vits14", 384),
    "b": ("dinov2_vitb14", 768),
    "l": ("dinov2_vitl14", 1024),
    "g": ("dinov2_vitg14", 1536),
}

dinov2_model, embed_dim = variant_config[variant]

# Load base DINOv2 model
model = torch.hub.load("facebookresearch/dinov2", dinov2_model)

# Download DinoBloom weights
ckpt_path = hf_hub_download(
    repo_id="virtual-human-chc/DinoBloom",
    filename=f"pytorch_model_{variant}.bin"
)
ckpt = torch.load(ckpt_path, map_location="cpu")

num_tokens = int(1 + (224 / 14) ** 2)
model.pos_embed = nn.Parameter(torch.zeros(1, num_tokens, embed_dim))
model.load_state_dict(ckpt, strict=True)
model.to(device)
model.eval()

# Get transforms
from torchvision import transforms
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]),
])

# Apply to image
from PIL import Image
img = Image.open("path/to/cell_image")
img_tensor = transform(img).unsqueeze(0).to(device)

# Get features
with torch.no_grad():
    features = model(img_tensor)

print(f"Features shape: {features.shape}")  # [1, 768] for DinoBloom-B

---

📊 Training Configuration

Model	Batch size	Train time	Feature dim	# Params
DinoBloom-S	1216	1:30 h	384	22 M
DinoBloom-B	960	0:45 h	768	86 M
DinoBloom-L	448	1:00 h	1024	304 M
DinoBloom-G	208	4:00 h	1536	1136 M

📊 Model Performance

DinoBloom outperforms existing medical and non-medical vision models in:

1. Linear probing and k-nearest neighbor evaluations for cell-type classification
2. Weakly supervised multiple-instance learning (MIL) for acute myeloid leukemia (AML) subtyping

Evaluation on peripheral blood: Image-level WBC classification on Acevedo dataset and patient-level AML subtyping on AML Hehr dataset. Performance is measured in weighted F1-score (wF1) and balanced Accuracy (bAcc).

Model / Dataset	Acevedo						AML Hehr
	1-NN wF1	1-NN bAcc	20-NN wF1	20-NN bAcc	Linear probe wF1	Linear probe bAcc	ABMIL wF1	ABMIL bAcc
DinoBloom-S	86.4	80.5	90.0	84.5	90.1	84.5	93.0	92.3
DinoBloom-B	87.4	81.9	90.5	85.4	90.7	85.5	92.7	91.9
DinoBloom-L	88.9	83.2	91.3	86.1	91.2	86.0	91.7	91.0
DinoBloom-G	89.1	83.5	91.4	86.4	91.8	86.6	93.1	92.4

Evaluation on bone marrow: WBC classification on the dataset BMC with 21 highly imbalanced classes.

	1-NN			20-NN			Linear probe
	wF1	Acc	bAcc	wF1	Acc	bAcc	wF1	Acc	bAcc
DinoBloom-S	78.4	78.3	62.0	84.2	84.8	55.6	85.7	85.9	71.4
DinoBloom-B	79.6	79.5	65.8	83.7	84.1	57.1	85.5	85.6	70.7
DinoBloom-L	78.8	78.8	57.7	83.6	84.0	56.3	84.9	85.0	64.4
DinoBloom-G	80.0	79.9	59.4	83.8	84.2	56.2	84.9	85.0	69.3

See the original paper for more details.

---

📖 Related Work

DinoBloom builds upon:

• DINOv2 - Self-supervised vision transformers
• Original DinoBloom Paper - MICCAI 2024

---

📄 Copyright

Code derived from https://github.com/marrlab/DinoBloom is licensed under the Apache 2.0 (See LICENSE file for details), Copyright (c) 2025 Ahmed Elnaggar. The ProtTrans pretrained models are released under the under terms of the Academic Free License v3.0 License, Copyright (c) 2025 Ahmed Elnaggar. The other code is licensed under the MIT license, Copyright (c) 2025 Maksim Pavlov.