Feature learning based classification and grading of diabetic retinopathy images

Abstract

Diabetic Retinopathy (DR) is one of the most common diseases that newlineaffects the vision of the eye. As a result of the prolonged high blood glucose newlinelevels that cause DR, the blood vessels within the retina can weaken and get newlinedamaged, resulting in vision problems and loss of vision. Manual examination newlineof retinal images is found to be tiresome and time-consuming. Nowadays, newlinethe research community has given many techniques for the early diagnosis newlineand detection of diabetic retinopathy. In this thesis, we explore approaches to newlineclassifying and grading the diabetic retinopathy images. newlineIdentifying the suitable features is significant for classification and newlinegrading of diabetic retinopathy images as the diabetic retinopathy images are newlinenormally hard to distinguish. This thesis aims to address two objectives: the first newlineobjective is to develop a model to detect the early signs of diabetic retinopathy . newlineThe second objective is to propose a hybrid framework that uses the generative newlinemodel-driven data representation with a discriminative model-based classifier. newlineWe propose a framework in which the features are extracted using newlinea convolutional neural network. We build class specific Gaussian mixture newlinemodels using training feature maps obtained from convolutional neural network. newlineFor each training and test feature map, we use the log likelihood scores newlineobtained by Gaussian mixture model as new feature representation respectively. newlineThus, we build a discriminative classifier such as support vector machine newlinein log likelihood vector machine. We name this model as hierarchical newlinemodel as CNN-GMM-SVM. We explore CNN-GMM-SVM for two tasks: newline1. Classification of normal and abnormal images. 2. Grading the severity newlineof diabetic retinopathy stages for the given retinal image. newline