Deep Learning with Heuristic Optimization Driven Diabetic Retinopathy Detection on Fundus Images

Abstract

Diabetic retinopathy (DR) is an illness occurred by the presence of diabetes which can resulted to blindness if left untreated. Identification of the DR at the benigning stage helps to prevent the loss of vision. Since deep learning (DL) models are commonly used for medical image analysis, it is used to classify the DR accurately. One of the effective way is to utilize a convolutional neural network (CNN) to classify retinal images as either normal or showing signs of DR. The CNN identifies the patterns and features in the images that are indicative of DR, such as the presence of microaneurysms, hemorrhages, exudates, or neovascularization. Therefore, this article presents an accurate DR grading and classification using Brain Storm Optimization with Deep Learning (DRGC-BSODL) algorithm. The DRGC-BSODL algorithm follows a three stage process. Initially, the contrast enhancement process is implemented. Next, the DRGC-BSODL model employs the BSO algorithm with multilevel thresholding (MLT) technique for image segmentation. Moreover, DenseNet169 model is exploited for generating a group of feature vectors. At the third stage, deep neural network (DNN) model is applied for DR classification. The simulation outcomes of the DRGC-BSODL model is tested on the fundus image dataset and the outcomes indicate the remarkable performance of the DRGC-BSODL model.