Optimized hybrid classifier for diagnosing diabetic retinopathy: Iterative blood vessel segmentation process

Abstract

AbstractIn general, diabetic retinopathy is a hurdle of diabetes that subsists throughout the world. Early detection of this severe disease through computer‐assisted diagnosis tools followed by the right treatment at the right time could control its terrible condition. From the last 2 years, numerous research efforts in this area have been introduced for the automatic detection of diabetic retinopathy with appropriate evaluations. However, there is a large variability in the databases and evaluation criteria used in the literature. Accordingly, this proposal tactics to develop a new contribution to automatic detection of diabetic retinopathy based on four main stages: “(a) image pre‐processing, (b) blood vessels segmentation, (c) feature extraction and dimension reduction, and (d) diabetic retinopathy recognition”. Two steps are used for accomplishing the image pre‐processing, (a) conversion of RGB into green channel image and (b) noise removal by median filtering. Further, the pre‐processed fundus image is subjected to Iterative segmentation‐based blood vessel segmentation. For performing the precise classification of the images, there is a prerequisite to extract the relevant informative features from the segmented blood vessels. Here, the features are extracted using discrete wavelet transform, and gray‐level co‐occurrence matrix. To attain the unique features with different information, the dimension reduction process is applied using principle component analysis. Finally, the Diabetic Retinopathy recognition is performed enabling a hybrid classifier, which merges the beneficial concepts of neural network, and convolutional neural network. As the main novelty, the number of hidden neurons in both neural network and convolutional neural network is optimized by the modified rider optimization algorithm called improvement counter‐based rider optimization algorithm intending to maximize the diagnostic accuracy. Moreover, convolutional neural network takes the transformed form of the segmented blood vessels using Discrete Wavelet Transform as input, and Neural Network takes dimension reduced features as input, and AND‐bit operation of the both classified outputs provides the diagnostic results, whether the corresponding image is normal or abnormal.