Accuracy Improvement in Diabetic Retinopathy Detection Using DLIA

G. Renith and A. Senthilselvi

In today’s modern technology, deep learning has become a strong giant technology based on neural network technique which imitates the human brain characteristics. There are many industries using deep learning technology including Health care, Automotive, Retail, Financial, and Oil & Gas and so on. Medical Imaging plays a vital role in Health Care industries. Medical imaging shows the interior structures of human body visually. Using current technology, we could able to diagnose different kind of diseases in the human body. Human Eye is the sensible organ that can receive visual images from outside world. There are many kinds of eye diseases including diabetes, glaucoma, cataract, AMD, hypertension, myopia and other abnormalities. Manual diagnosing process needs the person with good technical knowledge and more time. With deep learning approach, we could able to automatically diagnose the eye diseases with minimal time and effort. We collected some retinal color fundus images of both normal and diabetic images from ODIR 2019 database. Activation function plays a vital role in neural network technique. Activation function decides whether the information learned from the input data is relevant or not. It passes only the correct data information into the series of layers in the network. Hence activation function plays a main role in improving the accuracy in the diagnosis of eye disease. There are many activation functions in neural network including RELU, CELU, ELU, Tanh, SELU, Mish, Swish, Soft sign, Leaky RELU etc and we are going to use some of them. Another important method in improving the accuracy in disease diagnosis is image data augmentation. Augmentation is a technique used to increase the input dataset by modifying or varying little amount of input data from original images. We have increased accuracy by using Deep Learning Network with Image Augmentation (DLIA) technique by 3.3 % by our proposed method.

Volume 12 | Issue 4

Pages: 133-149

DOI: 10.5373/JARDCS/V12I4/20201426