Connectivity-Validated A/V Masks

• The paper presents an annotation pipeline that integrates AI initialization, manual correction, and interactive connected-component analysis to enforce continuous vessel connectivity.
• Connectivity-validated A/V masks accurately differentiate arteries, veins, and ambiguous segments via per-class layering, enhancing both clinical diagnostics and research.
• Quantitative validation using Cohen’s Kappa and Dice similarity demonstrates high inter-grader agreement and confirms the reproducibility of the segmentation approach.

Connectivity-validated artery/vein (A/V) masks are high-fidelity, pixel-level segmentations of retinal vasculature in color fundus images (CFIs), in which the anatomical connectivity of vessel trees is explicitly verified and corrected. These masks support research and clinical applications requiring topologically accurate separation of arteries, veins, and ambiguous or unknown vessel segments. The Rotterdam Artery-Vein (RAV) dataset introduced a systematic workflow for generating connectivity-validated A/V masks, employing AI-initialized masks, manual correction, per-class layering, and interactive connected-component analysis to ensure each vessel tree forms a single connected structure per class (Quiros et al., 19 Dec 2025).

1. Manual Annotation Pipeline

The generation of connectivity-validated A/V masks in the RAV dataset is based on a three-stage annotation pipeline:

1. Initial Mask Correction: Graders begin by loading an AI-generated binary vessel mask $M_0\colon \{1,\dots,1024\}^2 \to \{0,1\}$. Gross errors (false positives/negatives) are corrected using free-form tools to produce $M_1$, which represents all visible vessels.
2. A/V/Unknown Layer Assignment: Pixels in $M_1$ are classified into non-overlapping layers—artery ($A$), vein ($V$), or unknown ($U$):

$$A(x)+V(x)+U(x) = M_1(x), \;\; \forall\,x\in\{1,\dots,1024\}^2$$

Each channel is represented as a binary mask.

1. Manual Connectivity Verification: A connected-component visualization is activated for both artery and vein masks. Color coding highlights disjoint vessel fragments. Graders correct breaks and spurs to ensure each anatomical tree is a single connected component in the respective mask.

This annotation pipeline combines automated segmentation with layered manual tracing and emphasizes explicit, per-class connectivity integrity (Quiros et al., 19 Dec 2025).

2. Connected-Component Analysis and Correction

Connectivity checking is based on graph-theoretical analysis of binary masks:

• For any vessel mask $B\colon \mathbb{Z}^2 \to \{0,1\}$, a graph $G_B = (V_B, E_B)$ is defined, where nodes $V_B = \{p \mid B(p) = 1\}$ and edges $E_B$ connect neighboring vessel pixels.
• Neighborhoods are defined as either 4-connectivity ($N_4$) or 8-connectivity ($N_8$), with the latter including diagonal connections.

The correction process includes:

• Visualization: Each connected component is rendered with a unique color, revealing fragmentation.
• Manual Bridging: Disconnected fragments are joined by tracing minimal paths, setting $B(p) \leftarrow 1$ for each intervening pixel $p$.
• Morphological Operations (Manual): Small gaps are closed with brush tools approximating a morphological closing with a $3\times3$ structuring element.

No further algorithmic graph optimization is applied; all correction decisions are made interactively by annotators (Quiros et al., 19 Dec 2025).

3. Data Representations

Connectivity-validated A/V masks in the RAV dataset use a standardized PNG encoding:

Folder	Format	Channel Encoding
rgb/	8-bit PNG, RGB	Color fundus, circular FOV
contrast_enhanced/	8-bit PNG, RGB	Unsharp-masked vessel-enhanced
av_masks/	8-bit PNG, RGB	R: Artery, G: Unknown, B: Vein

Within av_masks/, each pixel encodes the segmentation as: $$(R,G,B)(x,y) = \begin{cases} (255,0,0) & \text{if } (x,y)\in A\ (0,0,255) & \text{if } (x,y)\in V\ (0,255,0) & \text{if } (x,y)\in U \end{cases}$$ This explicit channel separation allows unambiguous recovery of vessel class and supports automated pipeline integration (Quiros et al., 19 Dec 2025).

4. Quantitative Validation of Connectivity and Inter-Grader Agreement

No numerical before/after counts of connected components are reported, but inter-grader agreement provides indirect validation:

• A random window was used to measure Cohen’s Kappa ($\kappa$) and Dice similarity ($\mathrm{Dice}$) for the artery and vein masks.
  • Arteries: $\kappa_A = 0.882$, $\mathrm{Dice}_A = 0.906$
  • Veins: $\kappa_V = 0.890$, $\mathrm{Dice}_V = 0.899$

These agreement values indicate that the combination of automated initialization, manual layering, and connectivity correction yields highly reproducible A/V segmentation (Quiros et al., 19 Dec 2025).

5. Dataset Characteristics and Clinical Relevance

The RAV dataset provides 208 connectivity-validated fundus images spanning diverse acquisition modalities:

Aspect	Details
Total Images	208 (53 CC0, 155 under EULA)
Devices	8 distinct fundus cameras/OCTs (e.g., TOPCON, DRI OCT Triton)
Quality Variance	Deliberately broad (algorithmic quality 0.4–1.8, both “high” and “challenging”)
Field of View	Disc-centered, macula-centered, 35°/45°, left/right balanced
Demographics	Adults 40+ from the Rotterdam Study, varied ocular/systemic pathologies

The translational value is as follows:

• Topologically correct vessel trees enable quantitative retinal metrics (e.g., tortuosity, branch angles), which depend critically on uninterrupted connectivity.
• Machine learning models trained on this set inherit ground truth that enforces vascular structure, facilitating robust automated screening/diagnosis pipelines.
• Real-world variability in image quality and modality supports algorithm generalization to heterogeneous clinical data (Quiros et al., 19 Dec 2025).

6. Scope and Limitations

The workflow is notable for emphasizing manual, interactive quality control rather than fully automated validation or optimization. Connectivity is enforced via human-supervised bridging and deletion, with no automatic graph-based cost minimization. All ground truth reflects a consensus-focused process rather than single-expert annotation or algorithm-only outputs.

A plausible implication is that while reproducibility and connectivity are demonstrably high, total annotation time may be greater than alternatives lacking per-class connected component correction. However, the approach ensures each vessel class is a well-defined topological entity suitable for quantitative morphometric analyses and machine learning under challenging imaging conditions (Quiros et al., 19 Dec 2025).