Detection of Diabetic Retinopathy using Convolutional Neural Networks

Abstract

Diagnosing diabetic retinopathy (DR) from colour fundus images is a challenging and time-consuming task, requiring experienced clinicians to detect numerous small features and interpret a complex grading system. In our paper, we suggest using a Convolutional Neural Network (CNN) approach to automate DR diagnosis and accurately classify its severity from digital fundus images. Our developed CNN architecture, combined with data augmentation techniques, is capable of identifying intricate features crucial for classification, such as micro-aneurysms, exudates, and haemorrhages on the retina, enabling automatic diagnosis without manual intervention. We trained this network using a high-performance graphics processing unit (GPU) on the publicly available Kaggle dataset and achieved impressive results, especially for high-level classification tasks. In our experiments, utilizing a dataset comprising 3600 images, our proposed CNN attained an accuracy of 87% when validated against 500 additional images. These results demonstrate the effectiveness of our CNN approach in automating DR diagnosis with high accuracy.