An Automated OCT-Based Retinal Disease Classification Framework Using ShuffleNet and Enhanced Osprey Optimization

Abstract

In the world, the most common cause of visual impairment is retinal abnormalities. In the diagnosis of the retina, optical coherence tomography (OCT) is usually used, since it gives high-resolution cross-sectional images that show detailed retinal structures. In this research, a lightweight deep learning architecture called RetSEOS-Net is presented to use OCT images to recognize retinal diseases automatically. The framework initially uses a guided median filtering in minimizing OCT scans speckle noise. A model of attention-based U-Net is then employed to divide the retinal area, which is followed by a module based on ShuffleNet to obtain the efficient deep features. An Enhanced Osprey Optimization Algorithm (EOOA) is added to the features to refine them and work on the hyperparameters to increase the quality of features and the performance of the network. Finally, the fined features are categorized with a Softmax layer. The framework was tested on an OCT dataset that had four different categories namely diabetic retinopathy (DR), macular hole (MH) and central serous retinopathy (CSR) and normal. The recommended model achieved an accuracy of 96.79% in classification. The findings reveal that RetSEOS-Net is a good and computationally efficient model of automated screening of retinal diseases using OCT.