Binary Classification of Diabetic Retinopathy Using CNN Architecture

Abstract

     Diabetes mellitus (DM), a chronic, clinically heterogeneous condition, is becoming increasingly common all over the world. Insulin deficiency, resistance to insulin's actions on the body's surface, or both may lead to pancreatic beta-cell degeneration. Diabetes makes people more prone to its consequences, the most prevalent of which is diabetic retinopathy (DR). Deep learning algorithms surpass traditional classification models for illness diagnosis on photos of medical problems. Deep transfer learning models for medical DR detection were evaluated using the APTOS 2019 dataset. Deep transfer learning algorithms for medical diabetic retinopathy (DR) detection are being evaluated. MobileNet Convolutional Neural Networks (CNN) architecture is used to detect the DR in binary class classification tasks, which leverages pre-trained weights collected during the training process using the ImageNet database. Cohen Kappa, F1 score, recall, accuracy, and precision are some of the performance indicators used. According to the data, the given model is the most effective in terms of accuracy and training time for handling our challenges. Overall, MobileNet is a good pick. The following metrics were found to be accurate: 0.9455, precise: 0.94651, recall: 0.9455, F1 score:  0.94556, and Cohen Kappa score: 0.89083. This method might aid medical personnel in the early detection of diabetic retinopathy.