Rotterdam artery-vein segmentation (RAV) dataset

This paper introduces the Rotterdam Artery-Vein (RAV) dataset, a diverse collection of high-quality color fundus images with connectivity-validated artery-vein segmentation annotations derived from the Rotterdam Study, designed to support the development and benchmarking of robust machine learning models for retinal vascular analysis under real-world imaging conditions.

The Gist

Imagine your retina (the back of your eye) as a bustling city map. The blood vessels are the roads, and the doctors need to know exactly which roads are "arteries" (delivering fresh oxygen) and which are "veins" (taking away used blood). Why? Because the condition of these tiny roads tells a huge story about your overall health, from your heart to your brain.

For a long time, drawing these roads on a map was like doing it by hand with a pencil—slow, tiring, and hard to do for thousands of people. Then, computers got smart enough to do it automatically. But here's the problem: computers are only as smart as the examples they learn from. If you only show a computer maps of sunny, perfect cities, it will get confused when it sees a rainy, foggy city.

This paper introduces a new, massive "training school" for computers called the Rotterdam Artery-Vein (RAV) dataset. Here is the simple breakdown of what they did and why it matters:

1. The Problem: The "Perfect City" Bias

Most existing training data for eye scans is like a collection of photos taken only on perfect, sunny days with high-end cameras. They are clean, clear, and often focus on just one part of the eye (like the center of the city).

• The Reality: Real life is messy. People have cataracts, the cameras vary from old film models to new digital ones, and the lighting isn't always perfect.
• The Gap: If we train our AI only on "perfect" data, it fails when it meets a real patient with a slightly blurry photo or an older camera.

2. The Solution: The "Real-World" Training Ground

The researchers gathered 206 eye images from the famous "Rotterdam Study" (a long-term study of people in the Netherlands).

• The Mix: They didn't just pick the best photos. They intentionally included "messy" ones—images that other computer programs would usually throw away because they looked too blurry or dark.
• The Analogy: Think of it like teaching a driver. Instead of only letting them practice on an empty, sunny highway, this dataset throws them into rush hour, rain, fog, and different types of cars. This ensures the AI learns to drive in any condition.

3. The Secret Sauce: The "Co-Pilot" System

Drawing the lines for arteries and veins is incredibly hard. It's like trying to untangle a knot of red and blue yarn while blindfolded.

• The Old Way: Humans would stare at a photo and try to draw every single line from scratch. This takes forever.
• The New Way (RAV Method): The researchers used a smart computer to draw a rough outline of all the roads first. Then, human experts acted as "editors." They didn't start from zero; they just took that rough outline and colored the arteries Red and the veins Blue.
• The Result: This made the job much faster and allowed them to create a high-quality "answer key" for the computer to learn from. They even built a special tool that let the editors zoom in and fix any "broken roads" (gaps in the lines) to make sure the map was perfect.

4. The Treasure Chest

The result is a free, public library containing:

• The Photos: High-quality images of eyes (some clear, some challenging).
• The Answer Key: Digital masks that show exactly where the arteries and veins are, color-coded for the computer to read.
• The Metadata: Details about the patient's age, sex, and what kind of camera took the picture.

Why Should You Care?

This isn't just about better eye scans. It's about early warning systems.

• If a computer can accurately read these "road maps," it can spot tiny changes that signal high blood pressure, diabetes, or even early signs of Alzheimer's disease before a human doctor might notice.
• Because this dataset is so diverse (different ages, different cameras, different health issues), the AI trained on it will be more reliable when used in real hospitals around the world, not just in research labs.

In a nutshell: The authors built a massive, diverse, and "messy" library of eye maps with perfect labels. They did this to teach computers how to be better doctors, ensuring that when AI looks at your eyes, it understands the whole picture, not just the perfect parts.

Technical Summary

1. Problem Statement

Retinal vascular biomarkers (e.g., caliber, tortuosity, branching) are critical non-invasive indicators for systemic conditions such as hypertension, cardiovascular disease, stroke, and neurodegenerative disorders. However, the accurate measurement of these biomarkers relies on the precise segmentation and classification of retinal vessels into arteries and veins.

Current challenges include:

• Data Scarcity & Homogeneity: Existing publicly available Artery-Vein (A/V) segmentation datasets (e.g., RITE, DRIVE, HRF) are often limited in size, restricted to specific imaging devices, or focused on narrow patient demographics (e.g., only macula-centered or only diabetic retinopathy).
• Generalizability: Deep learning models trained on small, homogeneous datasets often fail to generalize to real-world clinical settings with diverse image qualities, acquisition devices, and pathologies.
• Annotation Cost: Manual annotation of A/V labels from scratch is labor-intensive and prone to inter-rater variability, hindering the creation of large-scale training sets.

2. Methodology

The authors developed the RAV dataset using a multi-stage pipeline designed to maximize diversity and annotation efficiency.

A. Data Sourcing and Sampling

• Source: Images were sampled from the Rotterdam Study (RS), a large prospective population-based cohort, along with the MYST and AMD-Life cohorts.
• Diversity: The dataset includes 206 Color Fundus Images (CFIs) capturing a wide spectrum of:
  • Devices: From older analog cameras to modern OCT machines.
  • Fields of View: Both macula-centered and optic disc-centered images.
  • Pathologies: A broad range of common ocular diseases and varying ages (40+).
  • Quality: Intentionally inclusive of "challenging" images (low visibility, artifacts) that automated quality assessment systems might reject, to ensure robustness.

B. Annotation Workflow (AI-Assisted Human-in-the-Loop)
To overcome the cost of manual labeling, the authors utilized a hybrid approach:

1. Initialization: A pre-trained vessel segmentation model generated an initial mask of the entire vascular tree.
2. Human Correction & Classification: Graders used a custom interface to:
   • Correct errors in the initial vessel mask.
   • Split the mask into three distinct layers: Arteries, Veins, and Unknown.
   • Topology Handling: In vessel crossings (where an artery crosses a vein), graders labeled both vessels on their respective masks rather than just the visible one. This preserves the underlying vascular topology for training.
3. Connectivity Verification: A specific tool visualized connected components, allowing graders to identify and fill gaps in vessel continuity, ensuring topological correctness.

C. Data Preprocessing

• All images were cropped to circular bounds and resized to a standardized 1024x1024 pixel PNG format.
• Three modalities are provided for each image:
  1. Original RGB fundus image.
  2. Contrast-enhanced version.
  3. RGB-encoded A/V Mask: Red channel = Arteries, Blue channel = Veins, Green channel = Unknown.

3. Key Contributions

• The RAV Dataset: A publicly available dataset of 206 high-quality, diverse CFIs with detailed A/V annotations. It addresses the gap in large-scale, heterogeneous datasets for retinal vascular analysis.
• Novel Annotation Protocol: A validated workflow combining AI-generated initial masks with human verification and explicit connectivity correction. This method significantly reduces annotation time while maintaining high quality.
• Topological Integrity: The specific instruction to label both vessels in crossing scenarios ensures that models trained on this data can learn the true 3D structure of the vascular tree, not just 2D projections.
• Real-World Robustness: By including low-quality images and diverse devices, the dataset is designed to train models that generalize well to clinical "messy" data.

4. Results and Validation

• Inter-Rater Agreement: Four graders annotated a subset of images to validate consistency. The results showed high agreement:
  • Arteries: Cohen's Kappa = 0.882; Mean Dice Score = 0.906.
  • Veins: Cohen's Kappa = 0.890; Mean Dice Score = 0.899.
  • Note: The authors acknowledge that agreement scores may be slightly overestimated compared to "from-scratch" labeling because the process started with a high-quality AI mask.
• Performance: The dataset was used to train the VascX system, which demonstrated state-of-the-art segmentation performance across various datasets, validating the utility of the RAV data.

5. Significance and Translational Relevance

• Clinical Impact: The RAV dataset enables the development of robust, generalizable machine learning tools for automated retinal screening. This can improve the early detection of systemic diseases (e.g., stroke risk, diabetes) and ocular conditions.
• Research Standard: By providing a diverse benchmark, it allows researchers to rigorously test algorithm performance under real-world variability (different devices, pathologies, and image qualities), moving beyond idealized lab conditions.
• Accessibility: The dataset is publicly available under a Creative Commons license, fostering open science and accelerating the development of ophthalmic AI.

Repository: The dataset is hosted on Dataverse NL (DOI: 10.34894/9OIMWY).