Effective Parameter Tuning of Convolution Network in Predicting Diabetic Retinopathy

Abstract

Diabetes is characterized by constant high level of blood glucose. The human body needs to maintain blood glucose at very narrow range. The patients who are all affected by diabetes for a long time affected by eye disease called Diabetic Retinopathy (DR). The retinal landmarks like Optic disc is detected and masked to reduce the false positive in the detection of Exudates. The abnormalities like Exudates, Microaneurysms and Hemorrhages are segmented to classify the various stages of DR. The proposed method is used to segment the retinal landmarks and retinal lesions for the classification of stages of DR using deep convolution network. DR is diagnosed and classified by ophthalmologist, optometrists and eye care professionals to determine whether the patients are in need of follow up or laser treatment. The screening of DR is performed using examination methods like slit lamp biomicroscopy, direct or indirect ophthalmoscopy and dilated or non-dilated digital fundus photography. The most common method used is by examining digital fundus photography. The manual screening process imposes a heavy workload on the ophthalmologists who have to evaluate a huge amount of fundus images every day. The other barriers to achieve recommended screening system are growing number of retinal disease affected patients and cost of current hospital-based system. The development of an automated screening system with image processing and deep learning techniques to diagnose DR is the potential solution to this problem. The growing number of diabetic patients has largely motivated the researchers in developing automated tools to facilitate the screening and evaluation procedures for DR. DR is the leading cause of blindness worldwide. A Decision Support System to assist ophthalmologists for diagnosing DR has been developed. Image processing algorithms are used to segment retinal landmarks like Optic Disc and Blood Vessels, and they are removed for further processing because they invariably appear in segmented output of white and red lesions. This screening system does not require any maintenance by doctors, patients or medical personals.