VAAR: Retinal Artery–Vein Dataset

• VAAR dataset is a comprehensive, manually annotated fundus image collection designed for precise pixel-level segmentation of retinal arteries and veins.
• It integrates diverse data sources and standardized annotation protocols, addressing limitations of smaller, heterogeneous prior datasets.
• VAAR underpins state-of-the-art models like VascX and LUNet, achieving high Dice scores and enhancing cardiovascular and ophthalmic disease analysis.

The Vessel ARtery–vein Annotation and Research (VAAR) dataset is an aggregation of manually annotated color fundus images (CFIs) specifically constructed for pixel-level segmentation of retinal arteries and veins. It underpins state-of-the-art model ensembles such as VascX (Quiros et al., 2024) and LUNet (Fhima et al., 2023), addressing the need for rigorously benchmarked, large-scale datasets with artery–vein differentiation and supporting robust automated microvasculature analysis. These resources are crucial for advancing research in automated identification of disease patterns, especially for cardiovascular and ophthalmic pathologies.

1. Origins and Motivation

The VAAR dataset was assembled from multiple public benchmarks as well as newly acquired and meticulously annotated CFIs from Dutch population-based cohorts, notably the Rotterdam Study. Its inception stemmed from limitations in prior datasets: previous artery–vein segmentation corpora (DRIVE, HRF-AV, LES-AV) contained fewer than 50 images each and were heterogeneous in resolution, field of view, and participant demographics, precluding reliable generalization for deep learning applications (Fhima et al., 2023). The VAAR collection thus integrates data diversity, quality stratification, and annotation protocol standardization to address these deficits.

2. Dataset Composition and Sources

VAAR comprises 562 CFIs, combining images from five public repositories and additional Dutch cohort studies. Each image is annotated at the pixel level for artery, vein, and “unknown” vessel classes.

Source	# CFIs	Field of View (°)	Annotation Method
RITE	40	Not specified	Manual, prior seed masking
HRF-AV	45	Not specified	Manual, prior seed masking
Les-AV	22	Not specified	Manual, prior seed masking
Leuven-Haifa	240	Not specified	Manual, prior seed masking
Rotterdam (ours)	215	30–50	Custom platform, pen tablet

All images were preprocessed to 1024×1024 px RGB PNGs (Quiros et al., 2024). For LUNet development, an independent subset (UZL Fundus VAAR, 240 images, 30° FOV, 1444×1444 px) was assembled via active-learning–guided, crowd-sourced annotation (Fhima et al., 2023).

3. Annotation Protocols

Annotations for VAAR utilized a vendor-independent desktop annotation platform. Workflow steps involved:

1. Initialization using AI-assisted binary vessel segmentation masks.
2. Manual labeling of each vessel pixel as artery, vein, or “unknown,” with ambiguous pixels consigned to the unknown layer.
3. Explicit marking of artery–vein crossings, duplicating intersection pixels onto both catheter layers.
4. Connected-components colorization to ensure contiguity and resolve branch gaps.

Four professional graders performed artery–vein annotation, with consensus training and periodic calibration meetings. For the UZL Fundus VAAR, annotation proceeded in two passes: an initial crowd-sourced phase by 15 medical students, followed by ophthalmologist review and correction (174/240 segmentations adjusted) (Fhima et al., 2023). Annotation reliability was not quantified by κ-statistics in the VascX corpus (Quiros et al., 2024).

4. Data Splitting, Stratification, and Metadata

VAAR supports stratified cross-validation and external benchmarking:

• 5-fold group cross-validation: Each fold consists of ~80% training, ~20% validation, grouped by eye to prevent data leakage.
• External testing: Out-of-distribution generalization assessed using held-out public benchmarks (e.g., HRF-AV, RITE).
• Stratification: Folds mix public and Rotterdam images and preserve eye-level grouping (Quiros et al., 2024).

For the UZL Fundus VAAR, stratified splits maintain patient, sex, and pathology balance; metadata includes patient_id, eye laterality, sex, age, clinical subgroup, device model, FOV, FundusQ-Net quality scores, and expert annotation status (Fhima et al., 2023).

5. File Formats and Directory Organization

All CFIs and associated masks are stored as PNG files:

File Type	Pixel Format	Channels
Fundus image	1024×1024 (or 1444×1444)	RGB
Artery mask	1024×1024 / 1444×1444	single-channel
Vein mask	1024×1024 / 1444×1444	single-channel
Unknown mask	1024×1024 / 1444×1444	single-channel
Crossing mask	1024×1024 (VascX)	single-channel

Directory structure hierarchically organizes images and masks into cross-validation folds and train/validation splits. For LUNet, all images are zero-padded to 1472×1472 before model input (Fhima et al., 2023).

6. Licensing, Access, and Availability

• Public components: RITE, HRF-AV, Les-AV, and Leuven-Haifa benchmarks are available under original terms.
• Rotterdam annotations: Not yet publicly downloadable; governed by Erasmus MC privacy regulations.
• LUNet VAAR resource: Will be open-access under CC-BY-4.0 pending publication, available on the ENRICH consortium website and GitHub (Fhima et al., 2023).
• VascX models and code: Available at https://github.com/EyeNED/VascX; raw Rotterdam CFIs excluded (Quiros et al., 2024).

7. Roles in Algorithm Development and Benchmarking

VAAR serves as the principal benchmark for segmentation models such as VascX and LUNet:

• Model training: Its breadth and diversity enable robust model generalization to varied patient demographics and imaging conditions.
• Performance evaluation: Quantitative metrics—Dice scores for arteries and veins—demonstrate superior performance for VascX and LUNet compared to previous methods.
  • LUNet achieves Diceₐ = 81.99, Dice_v = 84.54 on the test set (Fhima et al., 2023).
• Loss function protocols: Models leverage multi-channel losses with explicit unknown and crossing mask integration. LUNet employs continuity-aware loss incorporating $\|\nabla \hat{y}\|_1$ regularization and weighted sums of binary cross-entropy, Dice, and clDice terms.

A plausible implication is that the improved generalizability and segmentation fidelity traced to VAAR’s annotation scale and protocol support more accurate vascular feature extraction and downstream disease pattern recognition.

---

VAAR represents a consolidated, technically rigorous resource for retinal artery–vein segmentation, supporting both model development and disease research in ophthalmology and systemic microvascular analysis (Quiros et al., 2024, Fhima et al., 2023).