Grading of Diabetic Retinopathy using Different Machine Learning Algorithm

Abstract

Diabetes is a constantly recurring disease. It causes because, pancreas are not efficient enough to produce required amount of insulin.  One more reason is that insulin available within the body is not efficiently working. As of 2019, it is projected that around 450 million people had diabetes worldwide, with type 2 diabetes making up about 90% of the cases. Diabetic retinopathy (DR), also referred to as diabetic disease, may be a medical condition during which damage occurs to the retina and it’s a number one explanation for blindness. People who have  diabetes for more than 20 years, among those diabetic patient, 80% of them suffer from DR. Proposed method works with fundus image and processes it with optic disc detection, vessel segmentation, red lesion detection, and hard exudates evaluation, detection of Proliferative diabetic retinopathy (PDR) and grading of DR. For optic disc segmentation pre-processing method used is adaptive histogram equalization. For better classification, super pixels are generated using simple linear iterative clustering (SLIC) algorithm, and then k-means clustering algorithm is used as classifier. Proposed method is tested on database of 750 images and 98 % samples correctly extracted optic disc region. Blood vessel segmentation plays key role for detection of red lesion. Along with pre-processing method and feature extraction method, Gaussian mixture model is used as a classifier. Same database is used and gives accuracy of 93%. For red lesion detection optic disc region and wide and thick blood vessels are subtracted from original image and further given to Random forest classifier (RFC) for classification giving best result with accuracy of 96%, highest compared to other classifier such as support vector machine(SVM) 89%, k- means clustering 83%. Further hard exudates are evaluated using Convolution neural network (CNN) with accuracy of 99%.For detection of PDR, Matched Filter with Gaussian Kernel & Scale Elimination is used for feature extraction and for classification Thresholding is used based on Vessel Density (Vd) and Tortousity (Vt). Gradation of DR is characterized such as, Normal/Healthy-No lesion, Hard Exudates & Neovascularization present, Mild-0-5 Microaneurysms, Moderate- 5-10 number of  Microaneurysms , number of Hemorrhages >0, number of Exudates <=5, Severe- Lesions present are greater than that of moderate NPDR, PDR-Neovascularization Present. Further studies involve assessment with age factors, gender and samples from different living conditions. Use of proposed method will automatically detect and gives prior intimation regarding development of DR with gradation.