harvardairobotics/RNFLTNorm

Dataset Card: Color Fundus-to-RNFLT Norm Weights

Dataset Summary

Color Fundus-to-RNFLT Norm predicts a personalized expected-normal retinal nerve fiber layer thickness (RNFLT) map from a color fundus photograph. The predicted RNFLT norm is intended to capture patient-specific retinal anatomy and can be compared with an observed RNFLT map to derive a personalized RNFLT deviation map for glaucoma research.

Dataset Details

Dataset Description

Field	Value
Institution	Harvard Ophthalmology AI Lab, Schepens Eye Research Institute of Massachusetts Eye and Ear, Harvard Medical School
Task	Personalized RNFLT normative map prediction
Input	Color fundus photograph
Output	Predicted personalized expected-normal RNFLT map
Architecture	ResNet34-FPN residual predictor with a polar TSNIT head
Framework	PyTorch
Input resolution	384 × 384, plus a green-channel vessel high-pass
Output resolution	200 × 200
Model weights	checkpoint_best.pt (EMA weights, best validation epoch)
License	CC BY-NC-ND 4.0

• Authors: Elif Yildiz, Lucy Zha, Nazlee Zebardast, Min Shi, Mengyu Wang
• Contact: harvardophai@gmail.com, harvardairobotics@gmail.com

Model File

File	Description
checkpoint_best.pt	Pretrained color fundus → personalized RNFLT norm model (~364 MB)

Both files are required. The checkpoint stores the exponential-moving-average weights.

How to Download

Download from the Hugging Face website

Open Files and versions and download:

checkpoint_best.pt

Place the downloaded files in the checkpoint/ directory of the color-fundus code repository:

RNFLTNorm/
├── checkpoint/
│   ├── checkpoint_best.pt
│   └── mean_rnflt.npz
├── image/
├── rnfltnorm/
├── infer.py
├── inference.ipynb
├── requirements.txt
└── train.py

Download with Hugging Face Hub

pip install -U huggingface_hub

hf download harvardairobotics/ColorFundus2RNFLTNorm \
  --repo-type dataset \
  --local-dir checkpoint

Usage

The released implementation uses:

• Python 3.10
• PyTorch >= 2.0
• torchvision >= 0.15

From the code repository:

pip install -r requirements.txt
export PYTHONPATH="$PWD"

Load a sample color fundus image. The packaged sample PNGs are already cropped and in OD-like orientation, so pass crop_retina=False. Left-eye photographs that have not been mirrored should use flip=True.

from rnfltnorm.infer import (
    load_fundus_rgb,
    load_predictor,
    prepare_fundus,
    predict_rnflt,
    plot_2dmap,
)

rgb = load_fundus_rgb("image/fundus_1496028_1.png")

model, cfg, mean_np, device = load_predictor(
    "checkpoint/checkpoint_best.pt"
)

fundus = prepare_fundus(rgb, cfg, flip=False, crop_retina=False)
pred, mask = predict_rnflt(model, fundus, mean_np, cfg, device)
plot_2dmap(pred, mask, title="Predicted RNFLT Norm")

A complete example is provided in inference.ipynb.

Study Overview

The color-fundus model predicts patient-specific expected-normal RNFLT maps. Predictions can be compared with observed RNFLT maps to generate personalized RNFLT deviation maps and support research into glaucoma structure–function relationships.

OS photographs are mirrored so maps are OD-like (nasal to the right, superior up). Display thickness on a fixed 0–350 µm scale.

Uses

Direct Use

The pretrained weights are intended for research applications including:

• Reproducing the color-fundus RNFLT-norm model without retraining
• Predicting personalized expected-normal RNFLT maps from color fundus photographs
• Generating personalized RNFLT deviation maps
• Investigating glaucoma structure–function relationships
• Developing and evaluating personalized OCT normative modeling methods

Out-of-Scope Use

These weights are provided for research purposes only.

They are not intended for autonomous clinical diagnosis, clinical decision-making, treatment selection, or direct patient care. The associated study was released as a preprint and should be interpreted in that context.

Access

The weights are released under the CC BY-NC-ND 4.0 license for non-commercial research use.

The “Harvard” designation indicates that this research originated from investigators affiliated with the Department of Ophthalmology at Harvard Medical School and Massachusetts Eye and Ear. It does not imply endorsement, sponsorship, or assumption of responsibility by Harvard University or Harvard Medical School.

Citation

If you use the color-fundus code or pretrained weights, please cite:

@article{yildiz2026rnfltnorm,
  title={Deep Learning Prediction of Personalized Peripapillary Retinal Nerve Fiber Layer Thickness Norms from Fundus Images in Glaucoma},
  author={Yildiz, Elif and Zha, Lucy and Zebardast, Nazlee and Shi, Min and Wang, Mengyu},
  journal={medRxiv},
  year={2026},
  doi={10.64898/2026.05.26.26354081},
  url={https://pubmed.ncbi.nlm.nih.gov/42245038/}
}

Contact

For questions about the pretrained model or code, please contact:

• harvardophai@gmail.com
• harvardairobotics@gmail.com