Performance Analysis of Training Optimizers on Diabetic Retinopathy Detection Using SP-Net CNN

Diabetic retinopathy is an ocular disease linked to long-term diabetes mellitus and it can lead to permanent loss of vision if not addressed promptly. Timely detection of this disease is crucial as it can significantly reduce the risk of severe complications. Our aim is to identify the most effective optimization algorithm for training our deep learning (CNN) model. Our goal is to achieve high accuracy and efficiency in grading diabetic retinopathy into 5 different classes. In this research, we conducted a comparative evaluation of three optimizers namely Adam, SGD and RMSProp. For classification, we developed a convolutional neural network called SP-Net. An image based dataset named APTOS 2019 is taken from Kaggle and various pre-processing techniques are applied on it such as cropping, denoising, histogram equalization, and resizing. We used data augmentation techniques to tackle the class imbalance issue. Then the dataset was divided into three subsets (training, validation, and testing) using a split of 70:10:20. Our model SP-Net incorporates several layers including convolutional, dropout, max pooling, and fully connected layers. The results demonstrate that SP-Net achieved a highest test accuracy of 97.76% when it was trained using the SGD optimizer. Moreover, this model trained with the best optimizer outperformed the existing state-of-the-art techniques on comparison.