Detection of diabetic retinopathy and classification of its stages by using convolutional neural network

Abstract

Diabetes detection is pivotal in disease management and complication prevention. Traditional screening methods, like blood tests, are invasive and time-consuming. Deep learning has emerged as a non-invasive and automated alternative for diabetes detection. Convolutional neural networks (CNNs) excel in image analysis tasks, making them ideal for this purpose. This paper employs a CNN-based method for diabetes prediction using retinal images, utilizing the DenseNet169 architecture for feature extraction and diabetic retinopathy (DR) prediction. The APTOS 2019 blindness detection dataset from Kaggle, containing around 13,000 retinal images, is used for training. Pre-processing and normalization precede feature extraction, followed by the prediction of the DR stage. The model aims to classify retinal images into five stages of DR (0 to 4), ranging from no DR to proliferative DR. Our model achieved over 82% accuracy, outperforming advanced algorithms. Model evaluation includes accuracy, precision, recall, and F1 score measures.