RETFound OCT — MAE checkpoint (.pth)

This repository hosts RETFound_oct.pth: the full masked autoencoder (MAE) ViT-Large checkpoint used as the encoder in workflows that load RETFound with torch.load (custom ViT backbones, e.g. the RetinaPainter trainer). It is not the Hugging Face Transformers bundle (model.safetensors + AutoModel). Official RETFound code and context: rmaphoh/RETFound_MAE.

For Transformers / AutoModel / image-feature pipelines on the same research line, see iszt/RETFound_mae_natureOCT.

Access (gated)

This model uses gated user access with automatic approval.

1. Sign in at huggingface.co.
2. Open this model page and accept the access terms (contact details may be collected per Hub policy).
3. Create a read token: Settings → Access tokens.
4. Authenticate locally: huggingface-cli login or set the environment variable HF_TOKEN.

Until you complete step 2 for your account, programmatic downloads may return an access error.

Files

File	Description
RETFound_oct.pth	Full MAE checkpoint (encoder + decoder weights in one file). Load with PyTorch; downstream code often keeps only the ViT encoder weights.

Download (Python)

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="monish563/RETFOUND",
    filename="RETFound_oct.pth",
)
print(path)

Requires huggingface_hub and a logged-in session (or HF_TOKEN) after gated access is accepted.

Hugging Face CLI (alternative)

huggingface-cli download monish563/RETFOUND RETFound_oct.pth --local-dir ~/.cache/retina_painter

Use the same folder layout as above if your tool expects ~/.cache/retina_painter/RETFound_oct.pth.

RetinaPainter

In RetinaPainter RETFound modes, the trainer pulls this file via trainer/src/retfound_model.py into ~/.cache/retina_painter/RETFound_oct.pth. From the RetinaPainter repo root run python setup_retfound.py (interactive) or python setup_retfound.py --token YOUR_HF_TOKEN after accepting gated access on this page.

License

CC BY-NC 4.0 — non-commercial use with attribution. Commercial use is not permitted under this license unless you have separate permission from the rights holders.

Citation

If you use these weights or the RETFound methodology, cite the original paper:

Zhou, Y., et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023). https://doi.org/10.1038/s41586-023-06555-x

BibTeX:

@article{zhou2023foundation,
  title={A foundation model for generalizable disease detection from retinal images},
  author={Zhou, Yukun and Chia, Mark A and Wagner, Siegfried K and Ayhan, Murat S and Williamson, Dominic J and Struyven, Robbert R and Liu, Timing and Xu, Moucheng and Lozano, Mateo G and Woodward-Court, Peter and others},
  journal={Nature},
  volume={622},
  number={7981},
  pages={156--163},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

Disclaimer

These weights are for research and engineering (e.g. feature extraction, transfer learning). They are not validated for clinical decision-making and must not be used as a medical device without appropriate study, oversight, and regulatory compliance where applicable.