DIABETIC RETINOPATHY IMAGE CLASSIFICATION AND BLIND-NESS DETECTION USING DEEP LEARNING TECHNIQUES

Abstract

Diabetic retinopathy (DR) is a vision disorder that affects diabetic individuals. Timely detection and appropriate treatment of this condition are crucial to achieving favorable outcomes for patients. In recent times, Deep Learning (DL) applications have made significant advancements, finding widespread use in various healthcare and medical industries. In light of this, our research focused on exploring the efficacy of six different Convolutional Neural Networks (CNN) architectures for automated binary classification of diabetic retinopathy. The evaluated models were ResNet50, VGG16, DenseNet121, Inception-NetV3, AlexNet, and EfficientNetB3. Through pre-processing and modeling of these architectures, we assessed their performance using a range of metrics, including accuracy, Cohen kappa, macro avg., weighted avg., precision, recall, F1-score, and support. Our approach involved employing techniques such as confusion matrix, Epoch, and Keras model. The experiments were conducted using the APTOS dataset sourced from Kaggle, which is provided by the Asia Pacific Tele-Ophthalmology Society. Our study presents a comprehensive comparative analysis of the aforementioned CNN models. Among them, Insertion-NetV3 emerged as the most accurate, exhibiting a 0.73% accuracy rate and a Cohen kappa score of 0.779%. These findings contribute to the ongoing research in the domain of diabetic retinopathy detection using deep learning models.