Evolutionary Discriminative Deep Belief Network Based Diabetic Retinopathy Classification

The development of the glucose level in the blood may lead to vision threatening which is commonly known as Diabetic Retinopathy. Microaneurysms and Hemorrhages are the indications that illustrate the progressiveness of the eye disease named as Diabetic Retinopathy (DR). These indications appear as blood clots in the retinal area causing a serious threat to human vision. Digital fundus photographs from fundus cameras have the potential to explore the retinal area in a high resolution that helps to detect DR. Regular monitoring of fundus photographs helps diabetic patients from vision impairment. In this work, an Evolutionary Discriminative Deep Belief Network (DR-ED2BN) has been proposed to detect the Diabetic Retinopathy. Initially, the fundus photographs are preprocessed using Contrast Limited Adaptive Histogram based Equalization (ADHE) which enlightens the contrast and supports the noise removal in the image. The preprocessed image is subject to canny edge detection and Eigen value analysis to form clusters corresponding to the DR symptoms. The resultant image is processed to extract several statistical, shape and textural features that contribute to the feature set. These features further undergo Particle Swarm Optimization (PSO) to select the features predominantly indicating the DR symptoms. The reduced feature set is fed to Discriminative Deep Belief Network (D2BN) which classifies the fundus image as either DR healthy or DR affected. The proposed DR-ED2BN method works on real-world dataset and achieved accuracy, sensitivity and specificity of 89%, 92% and 65% respectively.