[Paper Review] Fully Automated Tree Topology Estimation and Artery-Vein Classification

Aashis Khanal, Saeid Motevali+1 more|arXiv (Cornell University)|Feb 4, 2022

Retinal Imaging and Analysis4 citations

TL;DR

This paper presents a fully automated, graph-based method for retinal vessel topology estimation and artery-vein classification from single-color fundus images. It combines deep learning-based vessel and AV probability maps with a novel cost function and high-level graph operations to correct topology errors, achieving state-of-the-art AV classification accuracy without manual intervention.

ABSTRACT

We present a fully automatic, graph-based technique for extracting the retinal vascular topology -- that is, how different vessels are connected to each other -- given a single color fundus image. Determining this connectivity is very challenging because vessels cross each other in a 2D image, obscuring their true paths. We quantitatively validated the usefulness of our extraction method by using it to achieve comparable state-of-the-art results in retinal artery-vein classification. Our proposed approach works as follows: We first segment the retinal vessels using our previously developed state-of-the-art segmentation method. Then, we estimate an initial graph from the extracted vessels and assign the most likely blood flow to each edge. We then use a handful of high-level operations (HLOs) to fix errors in the graph. These HLOs include detaching neighboring nodes, shifting the endpoints of an edge, and reversing the estimated blood flow direction for a branch. We use a novel cost function to find the optimal set of HLO operations for a given graph. Finally, we show that our extracted vascular structure is correct by propagating artery/vein labels along the branches. As our experiments show, our topology-based artery-vein labeling achieved state-of-the-art results on three datasets: DRIVE, AV-WIDE, and INSPIRE. We also performed several ablation studies to separately verify the importance of the segmentation and AV labeling steps of our proposed method. These ablation studies further confirmed that our graph extraction pipeline correctly models the underlying vascular anatomy.

Motivation & Objective

• To develop a fully automatic system for retinal vascular topology estimation from single fundus images, overcoming challenges posed by vessel crossings in 2D projections.
• To improve artery-vein classification accuracy by leveraging topological structure rather than relying solely on deep learning models.
• To create a general-purpose, explainable vascular feature extraction pipeline applicable to multiple retinal diseases.
• To validate the robustness and generalization of the method across diverse datasets with minimal hyperparameter tuning.
• To enable end-to-end, explainable diagnosis by integrating vessel segmentation, topology inference, and label propagation in a single framework.

Proposed method

• The method begins with two U-Net CNNs to generate vessel probability maps ($\hat{I}_p$) and AV likelihood maps ($\hat{I}_{av}$) from a fundus image ($I$).
• Multilevel skeletonization is applied to extract an initial union graph of the vasculature, followed by graph contraction to prune spurious components.
• Dijkstra’s shortest path algorithm is used to extract an undirected topology graph, which is then directed via a flow assignment algorithm based on vessel geometry and color.
• A novel cost function guides a sequence of high-level operations (HLOs)—including node detachment, edge endpoint shifting, and flow reversal—to minimize topology errors.
• The AV prior ($\hat{I}_{av}$) is mapped onto the directed graph ($g_d$), and a label propagation step assigns artery/vein labels along estimated branches.
• The cost function incorporates multiple factors: vessel likelihood ($w_v$), color similarity in Lab space ($w_c$), width-based weights ($w_w$), and a repel factor ($\Gamma$) to prevent overlapping vessel paths.

Experimental results

Research questions

• RQ1 Can a fully automated, graph-based method accurately reconstruct retinal vessel topology from a single 2D fundus image despite vessel crossings?
• RQ2 To what extent does incorporating topological structure improve artery-vein classification performance compared to deep learning models that operate on pixel-level features alone?
• RQ3 How robust is the proposed topology estimation pipeline across different datasets with varying image quality and vessel characteristics?
• RQ4 What is the contribution of each component—vessel segmentation, topology correction, and label propagation—to the final AV classification accuracy?
• RQ5 Can the extracted graph representation support the detection of clinically relevant features such as vessel breaks or pseudo-ONH?

Key findings

• The proposed method achieved state-of-the-art artery-vein classification accuracy on three benchmark datasets: DRIVE, AV-WIDE, and INSPIRE.
• The system generalized well to the INSPIRE dataset with minimal hyperparameter adjustment, indicating strong robustness and transferability.
• Ablation studies confirmed that improved vessel segmentation priors directly enhanced topology accuracy, validating the model's fidelity to vascular anatomy.
• The method outperformed semi-automated techniques with lower vessel recall, suggesting that topology-aware processing compensates for segmentation errors.
• The graph-based representation successfully identified potential vascular disruptions (e.g., breaks or obstructions), which may serve as diagnostic indicators.
• Label propagation along the estimated topology yielded consistent and accurate AV classification, demonstrating the utility of the extracted graph structure.