Retinal Hemorrhage Detection Using Supervised Learning A Hardware Implementation

Abstract

The human beings are gifted with visual perception of the surroundings with the important sensual organ, the eyes. Of the many medical conditions, diabetes is one of the leading reason for the loss of vision. Diabetes damages the human visual system and this pathology is referred to as diabetic retinopathy.The symptoms of diabetic retinopathy include retinal hemorrhages, microaneurysms, exudates and cotton wool spots. Early detection of the diabetic retinopathy symptoms will help the medical experts to prescribe suitable medication to prevent the patients from losing their eyesight. The work focusses on detecting retinal hemorrhages from fundus images for determining the possibility of diabetic retinopathy. Retinal hemorrhages are red lesions which appear on the retinal wall due to the leakage of blood vessels. Using different image processing techniques and machine learning, the early detection of these hemorrhages is proposed in the thesis. newlineInitially, retinal images are acquired from different sources such as publicly available database and also from a clinical database. The retinal images are split into training set and testing set. Before the image is processed for detecting DR symptoms, pre-processing is performed. Image enhancement techniques is utilised to remove the noise and balance the intensity distribution in the image. In image pre-processing, at first, the green channel extraction is performed. The red and the blue channels are affected more due to under illumination or over saturation. So the green channel was preferred over the other two. newlineDue to uneven illumination there can be vignetting artefacts on the periphery of the fundus image. These are removed using morphological erosion. The images can be noisy since the noise can interfere from external sources during capture of the image or from the image sensors. Two types of noise were normally found affecting the fundus image are the Gaussian noise and speckle noise. The noise elimination is performed using median filtering, were the median pi