Joint DR-DME classification using deep learning-CNN based modified grey-wolf optimizer with variable weights

Abstract

This paper aims to develop a computer-based diagnostic system to assist ophthalmologists in screening for diabetic retinopathy and diabetic macular edema by identifying the early signs of DR in retinal fundus images. The main objective of this work is to detect and grade DR by the severity of its classes using a hybrid deep-learning convolutional neural network-based modified grey-wolf optimizer with variable weights, which is termed as DLCNN-MGWO-VW. Initially, ResNet50 is used to extract the combined features of DR and DME diseases. Then, a disease-specific attention module is used to extract the disease-specific features, which differentiates the basic features of two considered diseases. Thesedata are later fed to the MGWO-VW for the optimal selection of individual disease-specific features from DR and DME, respectively. Finally, the disease-dependent attention module is used to identify the exact internal relationship between the two diseases with the aid of disease-dependent features that are also used for joint detection and classification of DR and DME, respectively. The simulations are performed on a publicly available IDRiD dataset, according to the ISBI-2018 challenge on DR segmentation and grading sub-challenge-2, where proposed hybrid DLCNN-MGWO-VW architecture maximizes the overall performance jointly for grading DR and DME,as compared to the state-of-art approaches from the literature. The proposed DLCNN-MGWO-VW methodoutperforms all the ISBI-2018 subchallenge-2 teams, with accuracy rates of 96.0%, 93.2%, and 92.23% for the detection and classification of DR, DME, and joint DR-DME,respectively.