Dataset Card: Fundus2RNFLT Weights

Dataset Summary

Pretrained model weights for Deriving OCT-Equivalent Retinal Nerve Fiber Layer Thickness Maps from Fundus Photographs with Deep Learning Improves Glaucoma Diagnosis (Shi et al., Ophthalmology Science 2026). The bundle contains the fundus-to-RNFLT U-Net, seven glaucoma classifiers reported in Table 8, and the RNFLT2Vec weights used for OCT artifact correction.

Code: Harvard-AI-and-Robotics-Lab/Fundus2RNFLT

Dataset Details

Dataset Description

Field	Value
Institution	Department of Ophthalmology, Harvard Medical School
Task	RNFLT map prediction from fundus photographs; glaucoma detection
Modalities	Color fundus photographs, OCT RNFLT maps
Scale	9 model checkpoints, 1.2 GB
License	CC BY-NC-ND 4.0

• Curated by: Lily Shi, Min Shi, In Young Chung, Louis R. Pasquale, Lucy Q. Shen, Mengyu Wang
• License: CC BY-NC-ND 4.0 — non-commercial research only
• Paper: Ophthalmology Science 2026
• Contact: harvardophai@gmail.com, harvardairobotics@gmail.com

Files

Path	Model	Used by
unet/unet-rnflt-efficientnet-b3/20250824_195258/	Fundus → RNFLT U-Net (EfficientNet-B3)	scripts/predict_rnflt.py --checkpoint weights/unet/.../model.pth
classifiers/fundus-resnet18/20250830_184641/	Fundus-only classifier	scripts/evaluate.py --runs-dir weights/classifiers
classifiers/rnflt_real-resnet18/20250831_201332/	OCT RNFLT-only classifier	same
classifiers/rnflt_pred-resnet18/20250907_172539/	Predicted RNFLT-only classifier	same
classifiers/fused_real-resnet18/20250907_120946/	Fundus + OCT RNFLT, concatenation	same
classifiers/fused_pred-resnet18/20250913_134801/	Fundus + predicted RNFLT, concatenation	same
classifiers/fused_real-resnet18_attn/20251022_221431/	Fundus + OCT RNFLT, attention fusion	same
classifiers/fused_pred-resnet18_attn/20251025_162943/	Fundus + predicted RNFLT, attention fusion	same
combined_rnflt2vec_weights_512_128_10_0001_004.93-0.03.h5	RNFLT2Vec artifact correction	scripts/correct_rnflt.py --weights weights/combined_rnflt2vec_weights_512_128_10_0001_004.93-0.03.h5

Each run directory holds model.pth (PyTorch state dict) and config.json. Classifier runs that read RNFLT maps also include rnflt_norm_stats.json. Keep the RNFLT2Vec .h5 file name unchanged: RNFLT2Vec parses the training epoch from it.

Verify integrity after download:

sha256sum -c SHA256SUMS

How to Download

From the Fundus2RNFLT repository root:

pip install huggingface_hub
huggingface-cli download harvardairobotics/Fundus2RNFLT --repo-type dataset --local-dir weights
sha256sum -c weights/SHA256SUMS

Uses

Direct Use

• Reproduce the pretrained models of Shi et al. (2026) without retraining
• Generate predicted RNFLT maps from fundus photographs
• Evaluate glaucoma classifiers on your own data prepared with the Fundus2RNFLT pipeline
• Artifact correction of OCT RNFLT maps with RNFLT2Vec

Out-of-Scope Use

Clinical decisions, patient care, or any commercial application. These weights shall not be used for clinical decisions at any time.

Access

The "Harvard" designation indicates these weights originate from the Department of Ophthalmology at Harvard Medical School. It does not imply endorsement, sponsorship, or assumption of responsibility by Harvard University or Harvard Medical School.

Citation

Fundus2RNFLT (BibTeX):

@article{shi2026fundus2rnflt,
  title={Deriving OCT-Equivalent Retinal Nerve Fiber Layer Thickness Maps from Fundus Photographs with Deep Learning Improves Glaucoma Diagnosis},
  author={Shi, Lily and Shi, Min and Chung, In Young and Pasquale, Louis R. and Shen, Lucy Q. and Wang, Mengyu},
  journal={Ophthalmology Science},
  volume={6},
  number={10},
  pages={101325},
  year={2026},
  publisher={Elsevier},
  doi={10.1016/j.xops.2026.101325}
}

APA:

Shi, L., Shi, M., Chung, I. Y., Pasquale, L. R., Shen, L. Q., & Wang, M. (2026). Deriving OCT-equivalent retinal nerve fiber layer thickness maps from fundus photographs with deep learning improves glaucoma diagnosis. Ophthalmology Science, 6(10), 101325. https://doi.org/10.1016/j.xops.2026.101325