A Deep Learning-based Transfer Learning Approach Fine-Tuned for Detecting Diabetic Retinopathy

Abstract

Diabetics have blood sugar levels that are abnormal. Because of this, their eyesight can be affected. Untreated diabetes has the potential to result in permanent blindness, a grave complication. Among the severe consequences, early detection stands out as a critical hurdle for ensuring effective treatment. Regrettably, accurately assessing fundus images relies on human expertise to determine the stage of diabetic retinopathy. A simplified detection process could greatly benefit a vast population. Leveraging convolutional neural networks (CNN), which have demonstrated efficacy across various domains, diabetic retinopathy identification has seen advancements. The possibility of preventing complete, irreversible blindness lies in timely identification, underscoring the need for a dependable screening mechanism. To enhance the efficiency of screening and reduce errors that plague current models, a novel system has been introduced, capable of categorizing distinct diabetic retinopathy stages. This study focuses on examining datasets of Indian diabetic retinopathy images, encompassing four classes: Microaneurysms (MA), Soft Exudates (SE), Hard Exudates (EX), and Hemorrhages (HE). By employing the proposed VGG16 architecture for feature extraction and the Logistic Regression classifier for classification, an impressive accuracy of 90.4% was achieved.