Hybrid Graph Representation Learning for Carotid Artery Stenosis Detection Based on Multimodal Retinal OCTA Images

Carotid artery stenosis (CAS) is one of the major causes of cerebral ischemic stroke. Rapid and precise detection of CAS is crucial for early intervention and reducing ischemic stroke incidence. Neuroimaging techniques, as the gold standard for evaluating cerebral abnormalities in CAS, suffer from limitations including expensive and time-consuming, hindering their use in large-scale screening. The ophthalmic artery is a branch of the internal carotid artery, several studies suggest that the biomarkers on retinal optical coherence tomography angiopraphy (OCTA) images are associated with CAS. Thus, retinal OCTA as a non-invasive and high-resolution imaging technique has potential as a suitable approach for identifying CAS patients. In this work, we developed a hybrid graph-based deep learning model to detect CAS from OCTA images. Given the differential impact of CAS on arteries and veins, we explicitly leverage the artery and vein information within the retinal region to enhance the sensitivity of the model to the change in microvasculature. We construct a hybrid graph representation by combining arterial and venous features, with the aim of improving the model’s ability to extract and integrate diverse anatomical information for more accurate CAS detection. For evaluation, we enrolled 182 CAS and 239 control subjects in this study. The experimental results demonstrated our retinal image analysis-based AI model, received promising results in distinguishing CAS and control subjects, with AUC of 0.7765 and an accuracy of 0.7750.