A Vessel-Aware Deep Learning Framework for Cardiovascular Risk Prediction using Retinal Images

Early detection of cardiovascular disease (CVD), a significant global health challenge, is crucial for preventing serious clinical issues. Retinal fundus imaging allows for non-invasive visualization of the microvascular system and shows great promise for automated assessment of cardiovascular risk. This paper introduces a deep learning framework that combines retinal vessel segmentation with fundus image analysis to predict cardiovascular risk based on Carotid Intima-Media Thickness (CIMT) measurements. A pretrained IterNet model generates vessel segmentation masks. These masks are combined with the original RGB fundus images to create a four-channel input representation. Three modern convolutional neural network architectures, EfficientNetB3, SE-ResNet50, and ConvNeXt-Tiny, are adjusted to handle this multimodal input and trained using Sharpness-Aware Minimization (SAM) optimizer. Experimental results show strong predictive performance in all assessed models, with EfficientNetB3 achieving the highest classification accuracy of 90.11 %. ConvNeXt-Tiny follows with 88.13 %, and SE-ResNet50 achieves 86.52 %. These findings demonstrate the value of including vascular structure in fundus-based analysis and suggest that the proposed method provides a scalable, accurate, and non-invasive option for early cardiovascular risk screening.