Exponential gannet firefly optimization algorithm enabled deep learning for diabetic retinopathy detection

Abstract

The diabetes complication which causes various damage to the human eye lead to complete blindness is called diabetic retinopathy. The investigation of the optimization-based Deep Learning (DL) approach is introduced for the detection of diabetic retinopathy using fundus images. Here, the fundus images are pre-processed initially using a median filter and Region of Interest (RoI) extraction, to remove the noise in the image. U-Net is used for lesion segmentation and trained using the introduced Gannet Pelican Optimization Algorithm (GPOA) to identify various types of lesions where GPOA is the integration of the Gannet Optimization Algorithm (GOA) and Pelican Optimization Algorithm (POA). The data augmentation process is carried out using flipping, rotation, shearing, cropping, and translation of fundus images, and the data-augmented fundus image is allowed for a feature extraction process where the image and vector-based features of fundus images are extracted. In addition, Deep Q Network (DQN) is used for the detection of diabetic retinopathy and is trained using the introduced Exponential Gannet Pelican Optimization Algorithm (EGFOA). The EGFOA is the combination of Exponentially Weighted Moving Average (EWMA), Gannet Optimization Algorithm (GOA), and Firefly Optimization Algorithm (FFA). Experimental outcomes achieved a maximum of 91.6% of accuracy, 92.2% of sensitivity, and 91.9% of specificity.