Diagnosis of Diabetic Retinopathy Using Deep Learning Approach Through Robust Data Security with Encrypted Image Analysis

Diabetic Retinopathy (DR) is a common complication of diabetes affecting the blood vessels in the retina and when left undiagnosed it precipitates in vision loss and blindness. DR is considered to be one of the principal causes of blind spells among adults all over the world. Diabetic retinal screening is normally done using trained personnel reviewing images acquired, which is tedious, tiresome, and has high variability in efficiency. This paper therefore seeks to design an automated classification system that can diagnose and classify high resolution retinal fundus images in an effort to indicate the severity of DR. The proposed system is trained using three Convolutional Neural Network (CNN) models: DenseNet121, InceptionV3, and Inception ResNetV2 to archive the most accurate results. Furthermore, a web application is created so that it could be easily deployed and is also coupled with strong data protection features incorporated within it. The proposed CNN-based model and its ensemble are proved to detect all stages of DR successfully. The experimental results indicate that our model achieves higher accuracy on the same dataset. It will have a great impact on the care of the patients as well as relieve the stress of the health professionals practicing in the diagnosis and management of DR.