iszt/RETFound_dinov2_meh

RETFound ViT-L/14 (DINOv2 → Transformers) — MEH AlzEye

Author of this fork: Dávid Isztl
Upstream project: RETFound_dinov2_meh by Yukun Zhou et al.
Paper: A foundation model for generalizable disease detection from retinal images, Nature (2023)

This repository provides a Transformers-compatible export of the RETFound DINOv2 encoder trained on a subset of MEH AlzEye (retinal CFP).
It includes config.json, model.safetensors, and an AutoImageProcessor, so you can load it directly with 🤗 AutoModel / AutoModelForImageClassification.

---

Model Details

Model Description

This is a ViT-Large/14 encoder pretrained with the DINOv2 objective on retinal color fundus photographs (CFP).
This fork converts the original PyTorch .pth checkpoint into a standard 🤗 Transformers format and removes DINOv2-only components.

• Developed by (upstream): Yukun Zhou et al.
• Shared by (this fork): Dávid Isztl
• Model type: Vision Transformer (encoder only)
• License: CC BY-NC 4.0 (inherited from upstream)
• Finetuned from: Upstream RETFound DINOv2 checkpoint (ViT-L/14)

Architecture (DINOv2 ViT-L/14 @ 224):

• hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, mlp_ratio=4
• patch_size=14, image_size=224, num_channels=3
• hidden_act="gelu", qkv_bias=True, layer_norm_eps=1e-6
• use_swiglu_ffn=False, use_mask_token=False

Conversion notes:

• Dropped DINOv2-only tensors: teacher components, momentum updates
• Remapped fused qkv weights (timm-style) → separate Q/K/V matrices (Transformers style)
• Set layer_norm_eps=1e-6 to match timm numerics
• Positional embeddings sized for 224×224 (patch 14×14)

Model Sources

• Repository (upstream): https://github.com/rmaphoh/RETFound
• Paper: https://www.nature.com/articles/s41586-023-06555-x

---

Uses

Direct Use

• Feature extraction from retinal images for downstream tasks
• Initial encoder for transfer learning on medical imaging research tasks (e.g., classification, retrieval)

Downstream Use

• Fine-tuning for image classification and related tasks using AutoModelForImageClassification
• Using CLS token or pooled features in custom pipelines

Out-of-Scope Use

• Clinical decision-making without proper validation and regulatory approval
• Commercial use beyond the CC BY-NC 4.0 license terms

---

Bias, Risks, and Limitations

• Trained on specific retinal data (subset of MEH AlzEye); distribution shifts (device, population, protocol) can degrade performance.
• Not a medical device; requires independent validation before any real-world or clinical deployment.
• Potential biases relate to dataset composition, imaging hardware, and labeling procedures.

Recommendations

• Perform task- and population-specific validation.
• Monitor for domain shift; consider domain adaptation where appropriate.
• Document preprocessing and augmentation pipelines for reproducibility.

---

How to Get Started with the Model

Feature extraction (encoder)

from transformers import AutoModel, AutoImageProcessor
from PIL import Image
import torch

repo = "iszt/RETFound_dinov2_meh"  # this fork

processor = AutoImageProcessor.from_pretrained(repo)
model = AutoModel.from_pretrained(repo)
model.eval()

img = Image.open("example_retina_cfp.jpg").convert("RGB")
inputs = processor(images=img, return_tensors="pt")

with torch.no_grad():
    out = model(**inputs)
    cls = out.last_hidden_state[:, 0]        # [B, 1024] — CLS embedding after final norm
    tokens = out.last_hidden_state[:, 1:, :] # [B, N, 1024] — patch tokens

Classification fine-tune (use AutoModelForImageClassification)

from transformers import AutoConfig, AutoImageProcessor, AutoModelForImageClassification

repo = "iszt/RETFound_dinov2_meh"
id2label = {0: "negative", 1: "positive"}  # example
label2id = {v: k for k, v in id2label.items()}

processor = AutoImageProcessor.from_pretrained(repo)

config = AutoConfig.from_pretrained(repo)
config.num_labels = len(id2label)
config.id2label = id2label
config.label2id = label2id

# Loads encoder weights from the repo and initializes a fresh classifier head
model = AutoModelForImageClassification.from_pretrained(
    repo,
    config=config,
    ignore_mismatched_sizes=True,  # replaces the classification head if shapes differ
)

# now train `model` with your dataloader/Trainer

---

Training Details

Training Data

• Upstream pretraining: retinal CFP from a portion of MEH AlzEye.

Training Procedure

• Objective: DINOv2 self-supervised pretraining.
• This fork: no additional training; checkpoint conversion only.

Preprocessing

• AutoImageProcessor provided for 224×224 inputs. If your dataset uses different normalization or resolution, adjust accordingly (and, if needed, interpolate positional embeddings).

Training Hyperparameters

• Not specified by upstream for this exact subset; see the paper and repository for general DINOv2 settings.

Speeds, Sizes, Times

• This fork only performs conversion; refer to upstream for compute details.

---

Evaluation

Testing Data, Factors & Metrics

• No new evaluation performed in this fork.
• For downstream tasks, report metrics relevant to the task (e.g., AUROC, accuracy, F1), and stratify by pertinent factors (device, demographics, pathology prevalence).

Results

• N/A for this fork; please cite/consult upstream results for baseline pretraining performance.

Summary

• Use this encoder as initialization; measure and report results on your target dataset.

---

Environmental Impact

This repository performs a format conversion only. Upstream pretraining compute and emissions are described in the paper and may be estimated via tools like the ML CO2 calculator.

• Hardware Type: N/A (conversion only)
• Hours used: N/A (conversion only)
• Cloud Provider / Region: N/A
• Carbon Emitted: N/A

---

Technical Specifications

Model Architecture and Objective

• Architecture: DINOv2 Vision Transformer Large, patch size 14, image size 224.
• Configuration: Uses Dinov2Config with standard GELU activation, no SwiGLU FFN, no mask token.
• Objective: DINOv2 self-supervised pretraining (encoder-only kept in this fork).
• Pooling: No pooling layer (use CLS token or custom pooling).

Compute Infrastructure

• This fork does not introduce new training; conversion was done locally.

Hardware

• N/A for conversion.

Software

• Conversion used PyTorch, timm, and 🤗 Transformers.

---

Citation

If you use this model, please cite the original RETFound paper:

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}
}

APA: Zhou, Y., Chia, M. A., Wagner, S. K., Ayhan, M. S., Williamson, D. J., Struyven, R. R., … et al. (2023). A foundation model for generalizable disease detection from retinal images. Nature, 622(7981), 156–163.

---

Glossary

• CFP: Color Fundus Photography
• DINOv2: Self-supervised Vision Transformer training method
• CLS token: Special token prepended to the patch sequence in ViT; often used as a global image representation.

---

More Information

• Upstream code and instructions: https://github.com/rmaphoh/RETFound
• Nature paper: https://www.nature.com/articles/s41586-023-06555-x

---

Model Card Authors

• Dávid Isztl (fork & conversion)

---

Model Card Contact

• For this fork/conversion: contact Dávid Isztl via Hugging Face.
• For upstream model/training code: ykzhoua@gmail.com or yukun.zhou.19@ucl.ac.uk.