Model card for LEMON

LEMON is an open-source foundation model for single-cell histology images. The model is a Vision Transformer (ViT-s/8) trained using self-supervised learning on a dataset of 10 million histology cell images sampled from 10,000 slides from TCGA. It is described in detail in its OpenReview paper.

LEMON can be used to extract robust features from single-cell histology images for various downstream applications, such as gene expression prediction or cell type classification.

How to use it to extract features.

The code below can be used to run inference. LEMON expects images of size 40x40 that were extracted at 0.25 microns per pixel (40X).

import torch
from pathlib import Path
from torchvision.transforms import ToPILImage
from model import prepare_transform, get_vit_feature_extractor

device = "cpu"
model_name = "vits8" 
target_cell_size = 40 
weight_path = Path("lemon.pth.tar")
stats_path = Path("mean_std.json")

# Model
transform = prepare_transform(stats_path, size=target_cell_size)
model = get_vit_feature_extractor(weight_path, model_name, img_size=target_cell_size)
model.eval()
model.to(device)

# Data
input = torch.rand(3, target_cell_size, target_cell_size)
input = ToPILImage()(input)

# Inference
with torch.autocast(device_type=device, dtype=torch.float16):
    with torch.inference_mode():
        features = model(transform(input).unsqueeze(0).to(device))

assert features.shape == (1, 384)

BibTeX entry and citation info.

If you find this repository useful, please consider citing our work:

@inproceedings{
  anonymous2025lemon,
  title={{LEMON} - a foundation model for single-cell nuclear morphologies for digital pathology},
  author={Anonymous},
  booktitle={Submitted to The Fourteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=JAalsmy7bZ},
  note={under review}
}