Early Diagnosis of Diabetic Retinopathy Through Exudate Detection Using Supervised Learning and Realization in Embedded System

Abstract

Diabetic retinopathy is caused by diabetes mellitus that damages the retinal wall of the eye. The medical condition gradually leads to vision impairment, if left untreated. The loss of vision can be prevented if diabetic retinopathy is diagnosed during the early stage of diabetes. The prevalent symptoms of diabetic retinopathy are microaneurysms, hemorrhages, exudates etc. Hard exudates are yellowish or white deposits, with a shiny tint, that can be seen in the retina. The exudates appear due to the leakage of lipids from the abnormal retinal capillaries of eye. Exudates can be diagnosed by an expert ophthalmologist. The assessment done manually cannot be completely relied upon, to suggest an effective remedial measure. This research study developed an automated non-invasive screening method for the diagnosis of hard exudates by processing the fundus images of patients. newline newlineTo conduct diabetic retinopathy screening the fundus images acquired are subjected to pre-processing procedures. Noise removal is achieved by using median filters where the noise pixel is replaced by the median value pixel. During the detection of exudates, it is mandatory to remove optic disk from the images since optic disc and exudates have similar characteristics. The shape of the optic disc is obtained by morphological operations which include dilation and erosion. newline newlineImage segmentation is performed by triangle thresholding method, Features are extracted using the marker image. The precise and non-redundant features are selected by expert ophthalmologists. The feature set is used for detecting the exudate pixels by training the Random Forest classifier and Support vector machine classifier. newline newlineRandom Forest classifier is basically a decision tree classifier. It generates a group of decision trees from the given random subset of training features which makes an ensemble. The votes obtained for individual decision trees decide the final class. Sometimes individual decision trees might be effected with noise but the aggregate of decision trees giv