Deep Learning-Based Multi-modal Algorithms for the Prediction of Diabetic Retinopathy

Diabetic retinopathy (DR) is a prevalent and debilitating condition that affects millions of people worldwide. It is an eye complication that can occur in individuals with diabetes, resulting from the impairment of retinal blood vessels. Timely detection and management of DR are pivotal to preventing vision loss and improving patient outcomes. Here, an algorithm based on deep learning for DR detection that uses only fundus photographs is proposed which trains and tests five deep learning models Convolutional Neural Network (CNN), ResNet, GoogleNet, InceptionV3, and VGG16, to classify the severity of DR. To assess the effectiveness of the suggested algorithms, a dataset consisting of 3662 fundus retinal images obtained from patients with and without diabetic retinopathy was utilized. The images were divided into five categories based on DR severity: No DR, mild, moderate, severe, and proliferate. The dataset is analyzed using the stated Deep Learning Algorithms to determine the optimum performance in terms of accuracy, training time, recall, prediction time, test score, precision, and the F1 score. The best algorithm is chosen after each one’s overall performance has been assessed. Experimental results show that the best algorithm achieves an accuracy of 97% in DR detection, which is higher than the accuracy achieved by other single-modal algorithms.