Diabetic Retinopathy Detection Using Deep Learning Multistage Training Method

Abstract

Diabetic retinopathy (DR) stands as the most prevalent diabetic eye ailment and constitutes one of the primary causes of blindness worldwide. Detecting and classifying retinal images can be laborious and demands specialized expertise. In this study, a convolutional neural network (CNN) was trained using stained retinal fundus images to identify DR and categorize its stages. The deep learning models chosen for this research encompassed InceptionResnetV2, VGG16, VGG19, DenseNet121, MobileNetV2, and EfficientNet2L. To enhance the resilience of the models and mitigate overfitting issues, data augmentation approaches were implemented. Each network underwent two levels of training. The initial level involved a feature extraction network with a customized classifier head, followed by fine-tuning the resulting network from the previous step through the unfreezing of certain layers. The efficacy of the proposed strategy was assessed through qualitative and quantitative evaluations using Kaggle’s diabetic retinopathy detection dataset. The obtained results demonstrated that our proposed methods, particularly those based on the refined InceptionResnetV2, achieved exceptional accuracy values, reaching 96.61%.