Deep Learning Models for Detecting Cardiovascular Disease from Retinal Images

Cardiovascular Disease (CVD) kills more people in the world than any other disease, and thus early disease detection is necessary for effective intervention. The retinal images are a non-invasive diagnostic approach where microvascular structures can be captured that are representative of cardiovascular health. This study compares four deep learning models (ConvNeXt based on ViT, EfficientNet, the customized Convolutional Neural Network (CNN)), and ResNet) to detect CVDs from retinal images. For training and evaluation, a large dataset of annotated retinal images was used and preprocessing steps were introduced to ensure consistency and to enhance feature extraction. Metrics such as accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC-ROC) were used to rate the models' performance. It is found that performance varies dramatically between the architectures and that the suitability of the architectures for clinical applications can be understood. An explainability component is also included to explain model predictions, to enhance clinical trust and understanding. Results corroborate the viability of measuring CVD using retinal images and present a systematic assessment of model effectiveness on this problem. By contributing to the development of reliable tools for CVD risk assessment using retinal screening, this study may be able to inform efforts toward creating novel, AI-assisted methods of early detection and treatment.