The FIREFLY Dataset

FIREFLY (Fundus Images REgistered with FLuorescein angiographY) is a retinal imaging dataset that pairs color fundus photographs with their corresponding fluorescein angiography (FA) captures, along with fine-scale vessel segmentation masks and artery/vein classification labels. The dataset is built by registering FA sequences to their fundus counterparts, aggregating vascular detail across FA frames using CNN-based methods, producing artery/vein labels via graph-based neural networks, and validating the final annotations with clinicians and imaging technicians.

A synthetic counterpart, FIREFLY-Gen, extends the dataset by generating realistic retinal fundus images together with matching vessel masks using diffusion and GAN-based generative models — expanding data availability for training and evaluating vessel analysis methods.

Request access to the FIREFLY-Gen dataset →

Why it matters

Standard fundus photography cannot reveal fine-scale vasculature, while FA does — at the cost of a contrast agent and a different appearance space. By registering the two modalities and distilling FA-level detail into fundus coordinates, FIREFLY enables supervision that was previously impossible: fundus-based models can now be trained and evaluated on vessel maps derived from FA-quality ground truth.

Construction pipeline

  1. Cross-modal registration between fundus and FA image pairs.
  2. Fine-scale vessel extraction from FA frames and projection into fundus coordinates.
  3. Artery/vein classification using hierarchical graph networks that leverage both modalities.
  4. Joint deep learning of fundus and FA for improved A/V classification.
  5. Synthetic augmentation (FIREFLY-Gen) using diffusion models to generate structurally realistic fundus images with paired vessel annotations.