Cardiovascular Risk Prediction From Retinal Images Using Vision Transformer

Abstract: Cardiovascular diseases are among the leading causes of death worldwide, making early diagnosis essential for effective treatment and prevention. This paper presents a deep learning-based approach for predicting heart disease risk using retinal fundus images and a Vision Transformer (ViT) model. The proposed system utilizes the RFMiD2 retinal image dataset, where images are pre-processed through resizing, normalization, and binary classification. Unlike traditional Convolutional Neural Networks (CNNs), the Vision Transformer divides images into patches and employs a self-attention mechanism to capture both local and global retinal features. Transfer learning with a pre-trained ViT model improves feature extraction and enhances performance on a limited medical image dataset. The model is trained using the Adam optimizer and Cross-Entropy Loss for binary classification. A Streamlit-based web application is developed to provide an interactive interface for uploading retinal images and generating predictions with confidence scores. The proposed model achieved an accuracy of 98.62%, demonstrating its effectiveness in identifying retinal patterns associated with cardiovascular risk. This framework offers a non-invasive, cost-effective, and AI-assisted screening solution that can support healthcare professionals in the early detection of heart disease and highlights the potential of Vision Transformers in medical image analysis.

Keywords: Heart Disease Prediction, Retinal Fundus Images, Vision Transformer (ViT), Deep Learning, Transfer Learning, Medical Image Classification, Artificial Intelligence, Early Disease Detection, Loss ,Accuracy.