Retinal Disease Classification Using Deep Learning

Abstract

Glaucoma, diabetic retinopathy and cataract are all retinal diseases that cause serious vision loss or blindness worldwide. Early diagnosis of these conditions is important to treat them effectively, and manual analysis of retina fundus images is time-consuming, with a possibility of human error. The aim of the present project is to develop an automated retinal disease classification system based on deep learning, namely, Convolutional Neural Networks (CNNs). We gathered a set of 4,272 fundus images that contain four categories, namely cataract, diabetic retinopathy, glaucoma, and normal. The data has been obtained in Kaggle and then pre-processed (resized, normalized, augmented, and pre-model performance). This was done with a standard machine learning workflow that trains, validates, and tests the prototype model. It scored 89 percent in all its classes. We calculated the performance based on evaluation metrics such as accuracy, precision, recall, and F1-score, and the confusion matrix confirmed that similar predictions were made within categories. The system has demonstrated that it can help reduce the workload of ophthalmologists, and help diagnose the disease early in under-resourced countries such as Bangladesh. Despite the limitations of the data used, the study identifies several opportunities in the future, such as extending the dataset, incorporating additional retinal conditions and employing newer deep learning architectures, such as ResNet50, Inception V3, VGG16 and VGG19. One day this research will end up with scalable, affordable and reliable diagnostic machines that can assist in increasing access to health care and improving vision.