Deep learning-based Retinopathy Diagnosis Using Fundus photography (FP) and Optical Coherence Tomography (OCT) images

Abstract

Diabetic retinopathy (DR) is a dangerous complication of diabetes that affects the retina and can lead to partial or complete vision loss. It progresses through five clinical stages, and the majority of patients are diagnosed at a late stage, when their vision is already severely impaired. Therefore, it is essential to diagnose DR in its early stage in order to prevent irreversible damage. With the rise of artificial intel ligence, deep learning has proven to be a useful tool in medical imaging, delivering accurate and fast diagnostic support. In this work, we explored the use of deep learning models for automatic detection and classification of diabetic retinopathy stages. We used two publicly available datasets: the APTOS 2019 dataset from Kaggle, which involves a five-class classification prob lem, and OLIVES, a multimodal dataset published by research institutions, used for binary classification between DR and Diabetic Macular Edema (DME).We tested different convolutional neural network (CNN) architectures , including custom-built and pre-trained models such as EfficientNet and InceptionResNetV2. To improve performance and address class imbalance,we applied training techniques like data augmentation, oversampling, and class weighting . Our results support that deep learning can effectively assist in the early diagnosis of diabetic retinopathy and en able the development of reliable and scalable screening systems in clinical practice.