Cardiovascular Disease Risk Assessment Through Retinal Image Analysis Using Deep Learning

Abstract

CVD is widely recognized to be the cause of death of almost 17.9 million people each year. This condition results in deaths of about one-third of all people [1]. As retinal vessels share the same developmental origin as coronary and cerebral microvasculature, fundus photography provides an unparalleled chance to assess the condition of cardiovascular system in a noninvasive manner. This study presents a model that allows to screen cardiovascular risk from retinal biomarkers in fundus images. The pipeline combines a clinically motivated preprocessing step — CLAHE contrast enhancement followed by Ben Graham local normalization — with three ImageNet-pretrained backbones (EfficientNetV2B3, InceptionV3, and ResNet50V2), each fine-tuned in two phases and combined through soft-voting of their sigmoid outputs. On the 724-image held-out test set, the ensemble achieved 96.4% accuracy, 0.987 AUC-ROC, 97.9% recall, and 93.5% specificity at the Youden-optimal threshold of 0.5127. The metrics are comparable to the IEEE CSNT 2025 baseline, although the data distributions and evaluation schemes were different in the two studies. Grad-CAM heatmaps allow spatial inspection of model attention, and Test Time Augmentation together with Youden’s J threshold optimization improve reliability under realistic screening conditions. All experiments draw on a composite dataset of 7,223 fundus images pulled from eight of the nine sources used in the baseline study; we dropped one source over resolution incompatibility.