Enhanced Graph Representation for Reliable Retinal Vessel Analysis: CLSA Population-Based Study

Several fully automatic pipelines have recently been published to calculate vascular biomarkers on fundus images, relying on deep learning models for the segmentation of the arterioles and venules (A/V). But the artefacts produced by those probabilistic models can be problematic for vascular graph parsing algorithms. We propose a graph-based heuristic-driven method to estimates the most plausible (according to clinical knowledge) vascular graph given an imperfect A/V segmentation map. Experiments show significant improvements in the classification of arterioles and venules on the MAPLES-DR dataset and enhance the reliability of vessel tortuosity measured on the CLSA dataset.