A GRAPH-DRIVEN HYBRID CNN-GNN APPROACH FOR VESSEL CLASSIFICATION IN ADAPTIVE OPTICS IMAGES

Abir Aissa, Guillaume Lachaud, Florence ROSSANT, Helene Urien, Michel Paques

Abstract:

Adaptive Optics Ophthalmoscopy (AOO) enables high-resolution retinal imaging for clinical biomarker analysis.
While existing methods automatically segment retinal bifurcations and extract vascular biomarkers, they require manual selection of high-quality arterial regions, limiting scalability
and efficiency. We propose a fully automated patch-based
CNN-GNN framework for vessel classification in AOO images. Our method extracts deep features from image patches,
constructs graphs with cost-weighted edges capturing vessel
connectivity, then applies Graph Attention Networks (GATs)
for classification. This hybrid approach integrates local features with global topological information through graph
processing, eliminating manual region selection requirements. We evaluate our pipeline on a dataset of 4,258 vessel
patches across four CNN/Transformer backbones (ResNet-
18, EfficientNet-B2, DenseNet-121, TinyViT). The backbone+GAT combination consistently outperforms backbone-only baselines, achieving up to 86.5% accuracy with improvements ranging from 1.9% to 4.4% across architectures.
Our approach provides a robust, scalable foundation for
automated vascular biomarker analysis in AOO imaging,
supporting more efficient clinical workflows.