Diabetic Retinopathy Detection using MobileNetV2 Architecture
Abhay M Pamadi, Ananya Ravishankar, P Anu Nithya, Gorantla. Jahnavi, Sheela Kathavate
Abstract
The disease Diabetic Retinopathy (DR) is a microvascular diabetic condition that affects the eyes. It is attributed to the impairment of the retinal blood vessels. The later it is detected, the greater the likelihood that the patient will lose sight. This paper proposes two Convolutional Neural Network (CNN) models, one of them a binary classification to detect retinopathy and another multinomial classification model to further classify retinopathy into five distinct and widely used stages - None, Mild, Moderate, Severe and Proliferative DR. Using Gaussian filtered fundus images enhances the recognition of subtle features such as edges or spots used for diagnosis. Transfer learning on a pre-trained MobileNetV2 model further enhances the accuracy to 78% for a multinomial classification and up to 97% for binomial classification.