Diabetic retinopathy detection from image to classification using deep convolutional neural network

Abstract

Purpose : Diabetic Retinopathy (DR) is an eye disease associated with long-standing diabetes that is one of the leading causes of blindness worldwide. Classification of DR severity is important for disease management and research. We developed a fully automated supervised deep convolutional neural network (CNN) to detect DR severity from fundus retinal images.

Methods : The proposed method classifies 2D fundus image as Normal, Mild DR, Moderate DR, Severe DR and Proliferative DR. The database of 35126 high-resolution color retinal fundus images is provided by EyePACS and made publicly available by California Healthcare Foundation at https://www.kaggle.com . The images were taken under a variety of imaging conditions with different models of cameras, magnification, and image quality. Ninety percent of images were selected for training and the rest for testing. The classes had uneven image distributions as shown in Fig.1.a, that biases the deep CNN training toward Normal class. The different image colors and illumination made it hard for deep CNN to converge. Preprocessing to create image consistency and brightness normalization were used to solve the problems. First images were normalized with respect to the standard reference image . Then images of training set were augmented by rotations (multiple of 10 degree). The augmented images were randomly selected from different classes with respect to their relative size to build new training set including images with equally distributed classes. Finally images were sub-sampled and cropped into size 512x512 to reduce computational cost of training. Some of original images and their corresponding cropped normalized in different orientations are shown in Fig1.b and c respectively. A deep CNN with 10 convolutional Layers and 5 pooling layers and 3 fully connected layers (Fig.2.a) was used to classify the images into five levels of severity as normal, mild, moderate, severe, proliferative.

Results : The best performance of the deep CNN was achieved with 200k iteration (around 5 epochs ) with overall accuracy of 75.6%. as shown in Fig.2.b. The performance is improving with more epochs.

Conclusions : Deep CNN successfully classified images of diabetic retinopathy on a challenging dataset into levels of severity in high speed (batch processing) and without human intervention, using supervised learning that shows efficiency of deep CNN on processing Big data.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.