Detection and Grading of Diabetic Retinopathy using Textural Feature Descriptors and Machine Learning Algorithms

Abstract

Diabetes mellitus (DM) is a metabolic condition characterized by high blood glucose newlinelevels, leading to damage of the vital organs of the human body. Diabetic retinopathy newline(DR), a severe complication of DM, is a microvascular disease of the eye that can cause newlineirreversible blindness. DR can lead to vision loss and blindness if left untreated at an early stage. Therefore it is crucial for diabetic patients to undergo frequent eye screening for newlineearly recognition and treatment. Since manual screening and detection are time-consuming newlineand subjective, automated DR grading systems are necessary for timely treatment. Digital newlinefundus images of the retina serve as a screening platform to diagnose DR and grade the newlinevision problems. The task of accurately grading DR in an automated manner is challenging. newlineAdvancements in image processing and machine learning algorithms lead to effective ways newlinefor fast and accurate diagnosis. The research intends to develop automated DR grading newlinealgorithms using machine learning based on feature descriptors by advanced medical image newlineprocessing techniques. newlineSeveral novel algorithms for categorizing DR fundus images using leading textural newlinefeatures and machine learning are presented in the thesis. The research initially focuses newlineon the early detection of DR using a transform-statistical feature extraction technique. It extends to a multi-class DR grading system using textural features based on multiresolution newlinemicro-macro feature descriptors. Pre-trained convolutional neural network (CNN) models are employed further for grading of diabetic retinopathy. A more accurate DR grading newlinesystem is obtained using an ensemble of multi-stage deep convolutional neural networks. newlineThesis also proposes hierarchical clustering by Siamese network (HCSN) with pre-trained newlineCNN models for DR grading. newlineEarly detection of the pre-eminent indicators of DR, called microaneurysms, using newlinediscrete orthonormal Stock well transform (DOST) and statistical features is proposed. newlineSkewness and kurtosis of local binary pattern (LBP)