DiscFoundGlobal

A DINOv2 ViT-Small model further pre-trained on over half a million optic disc crops sourced from public datsets from 5 continents trained via self-supervised token reconstruction. Model can be frozen or finetuned for multi-purpose downstream tasks.

Attention maps generated by DiscFoundGlobal for predicting OCT-derived structural parameters (BMO Area, MRW sectors, RNFL sectors) from optic disc fundus photographs. Brighter regions indicate higher attention weight. Note that the model learned to identify the anatomical regions correlating to the measured OCT prameters.

---

Download DiscFoundGlobal.pth from https://github.com/stephanmkonig/DiscFoundGlobal/releases/tag/Model.

Loading the model

import timm
import torch
model = timm.create_model('vit_small_patch14_reg4_dinov2', img_size=(392, 392), num_classes=0)
state_dict = torch.load('DiscFoundGlobal.pth', weights_only=True)
model.load_state_dict(state_dict)
model.eval()

Inference

Images must be resized to 392×392 and normalized with mean=0.5, std=0.5.

import torch
from torchvision.transforms import v2 as T
from torchvision.io import read_image
transform = T.Compose([
    T.Resize((392, 392), antialias=True),
    T.ToDtype(torch.float32, scale=True),
    T.Normalize(mean=[0.5], std=[0.5]),
])
img = read_image('fundus.jpg')           # [C, H, W] uint8
img = transform(img).unsqueeze(0)        # [1, 3, 392, 392]
with torch.inference_mode():
    tokens = encoder.forward_features(img)   # [1, 789, 384]

forward_features returns 789 tokens: 1 CLS token, 4 register tokens, and 784 patch tokens (28×28 patches).