Classification of Glaucoma from Retinal Fundus Images using Machine Learning Algorithms

Abstract

Medical image processing plays a vital role in detecting and diagnosing numerous newlinediseases. Among those, glaucoma is one leading ophthalmic disease that occurs due to a rise in the Intra Ocular Pressure (IOP) inside the human eye and may lead to peripheral vision loss. Glaucoma is commonly referred to as a sneak thief of sight because of its invisible symptoms in the early stages. It is necessary to detect glaucoma in the early stages; otherwise, it may lead to blindness. Many researchers have proposed various methods for detection using medical and Computer Aided Design (CAD) approaches;however, there are a few gaps concerning variations in retinal image characteristics. newlineTherefore, in this dissertation, three CAD-based frameworks are reported for the classification of glaucoma from retinal fundus images. newlineThe first framework reports a new methodology based on pattern descriptors and newlinestatistical texture features. In the first phase, anisotropic filtering is performed on the input images for noise reduction without blurring the edges. Next, a Local Directional newlineTexture Pattern (LDTP) is applied to encode the image information into the pattern and newlinethen extract the first-order and second-order statistical texture features to examine the newlinespatial relation between the pixels. The given set of images is classified using the feature vector and then fed to Support Vector Machines (SVM) and decision tree algorithms for classifying healthy or glaucoma images. newlineIn the second framework, shape and texture features are combined to classify glaucoma newlinefrom given fundus images. Initially, the input images are processed in local regions by applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) technique. newlineLater, significant sub-band images are obtained with the Quasi Bi-variate Variational Mode Decomposition (QB-VMD) method. From the sub-band images, shape information is obtained using a Pyramidal Histogram of Oriented Gradients (PHOG), and texture information is obtained using invariant Haralick features. Finally, ima