Classification of Diabetic Retinopathy Images Based on Customised CNN Architecture

Diabetes is a disease emerging to be a big threat to humanity, which even after such scientific and medical advance-ment is still incurable. Its only remedy is early detection and precautionary measure to reduce its effects to minimum. Since it affects all parts of body parts and organs hence there are ways to detect its presence before it critically damages the body. Eyes retina is also affected by diabetes, causing blood vessels in the retina to rupture and due to some complication eventually causing permanent blindness. Luckily, we can take images of retina using retinopathy. These images can be utilize to detect Diabetic Retinopathy. This paper implements automated tools to detect Diabetic Retinopathy using these images. The paper uses CNN approach for the classification of DR images. We have used pre-trained CNN models i.e. AlexNet, VGG-16 and SqueezeNet, which gave the classification accuracy of 93.46%, 91.82% and 94.49% respectively. Also, a customized 5 layered CNN model is proposed which consists of 2 convolution layers and 3 fully connected neural layers, this methodology has shown promising result of sensitivity, specificity and accuracy with numbers of 98.94%, 97.87% and 98.15% respectively.