STATISTICAL ANALYSIS OF RETINAL ARTERIAL NETWORKS: Graph-Based Radiomics Framework

Traditional radiomics enables mathematical and statistical analysis of
medical images but remains limited in capturing the structural organization and complexity of vascular networks. This study analyzes the retinal arterial vasculature in relation to Cardiovascular Disease Risk (CVDR), using a graph-based radiomics framework, built on adjacency matrices derived from vascular graph representations. Images with near visible vasculature were selected from the Retinal Fundus Multi-disease Image Dataset (RFMiD) and Ocular Disease Intelligent Recognition (ODIR) dataset. Full MESSIDOR dataset was used reference. Shannon entropy values of 1.4−3.2 bits, supported by Pearson correlation analysis, indicate substantial feature variability and suitability for machine learning (ML) applications. Six configurations (a gradient-boosted decision tree baseline, a metrics-only MLP, GAT and GCN alone, and each fused with the calculated metrics) are evaluated with respect to their discrimination performance. To improve model interpretability, explainable artificial intelligence (XAI) techniques, specifically SHAP (SHapley Additive exPlanations) and GNNExplainer, are employed to quantify the contribution of individual network metrics to model predictions. Results reveal the following contribution pattern: high Maximum degree consistently drives predictions toward higher CVDR risk, and high Diameter toward lower risk in RFMiD and
MESSIDOR, whereas Global efficiency and Global clustering coefficient reverse direction across datasets. Only MESSIDOR shows a measurable signal, with smaller Diameter and higher Modularity in CVDR images and above-chance discrimination (AUC up to 0.62), while RFMiD and ODIR remain near chance.