Image Analysis of Retinal Vascular Network Geometry and its Relationship to Cardiovascular Complications

Objective: 1) To detect retina characteristics that associated with stroke; 2) To develop a statistics model with variables of retina characteristics for classifying patients with stroke from those without stroke in aged population.;Method: Matched case control study. Patients with stroke from the diabetic retinopathy screening program and stroke patients from Acute Stroke Unit were selected as stroke cases. Controls (patients without history of stroke) with matched diabetes status and age were selected from the diabetic retinopathy screening program and eye outpatient clinics. All subjects in this study were from Prince of Wales Hospital, Hong Kong. Risk factors of stroke from all subjects were collected, including age, gender, diabetes, hypertension, hyperlipidemia, history of ischemic heart disease, atrial fibrillation and smoking. Color retina images of each subject were collected and analyzed. The retina characteristics, including diameters of arterioles and venules, bifurcation coefficients, bifurcation angles, branch symmetry, optic disc perimeter were extracted from the color retina images by software "ImageJ". Other retina characteristics including arteriole-venule nicking, hemorrhages, exudates, arteriole occlusion, and vessel tortuosity were also recorded. Independent t test and Chi-squire test were used to compare the continuous and categorical retina characteristics respectively between patients with stroke and those without stroke. Logistic model combining the risk factors of stroke and retina characteristics was established to classify patients with stroke from those without stroke. All data analysis was by SPSS 16.0.;Results: there were 122 stroke cases and 122 controls recruited in this study. There were 41 patients without diabetes and 81 patients with diabetes in each group. Retina characteristics including diameters of arterioles and venules, vessel tortuosity, hemorrhages, exudates, arteriole-venule nicking were significantly different between the two groups. We established risk models to classify patients with stroke from those without stroke. The risk model with highest accuracy of classification included 1) stroke risk factors including hypertension, diabetes and atrial fibrillation; 2) retina characteristics, including arteriole diameters, vessel tortuosity, hemorrhages, arteriole-venule nicking and venule symmetry; 3)interaction between retina characteristics, including arteriole diameters by venule symmetry, arteriole diameters by hemorrhage, and venule symmetry by vessel tortuosity. The accuracy of classification was 80.4%. Using retinal characteristics alone achieved an accuracy of 74.5%.;Conclusion: color retina images are a potential tool for stroke risk stratification. Useful characteristics found in the retinal images included vessel diameters, vessel tortuosity, vessel symmetry, hemorrhage, arteriole-venule nicking. The association between the retinal characteristic and stroke was modified by other retinal characteristics.