Detection of central serous retinopathy using deep learning through retinal images

Abstract

The human eye is responsible for the visual reorganization of objects in the environment. The eye is divided into different layers and front/back areas; however, the most important part is the retina, responsible for capturing light and generating electrical impulses for further processing in the brain. Several manual and automated methods have been proposed to detect retinal diseases, though these techniques are time-consuming, inefficient, and unpleasant for patients. This research proposes a deep learning-based CSR detection employing two imaging techniques: OCT and fundus photography. These input images are manually augmented before classification, followed by training of DarkNet and DenseNet networks through both datasets. Moreover, pre-trained DarkNet and DenseNet classifiers are modified according to the need. Finally, the performance of both networks on their datasets is compared using evaluation parameters. After several experiments, the best accuracy of 99.78%, the sensitivity of 99.6%, specificity of 100%, and the F1 score of 99.52% were achieved through OCT images using the DenseNet network. The experimental results demonstrate that the proposed model is effective and efficient for CSR detection using the OCT dataset and suitable for deployment in clinical applications.