The role of deep learning-based algorithms in retinal image processing: a review

The use of deep learning techniques in retinal image processing has revolutionized the early diagnosis of eye diseases. Particularly, deep learning-based methods are increasingly being used to detect significant anatomical structures of the retina, such as the optic disk, optic cup, macula, fovea, and vessels, as well as for the recognition of various eye diseases. This review article discusses various applications of deep learning techniques on fundus retinal images and summarizes recent developments in this field. The article covers many topics, including lesion detection and segmentation, biomarker segmentation, disease diagnosis, image synthesis, and other applications. The section on lesion detection and segmentation details various CNN, FCN, U-Net, and Mask-RCNN architectures used to detect specific lesion types such as hemorrhages, microaneurysms, exudates, and drusen. The biomarker segmentation section focuses on vessel and optic disk/cup (OD/OC) segmentation, examining innovative methods such as encoder-decoder architectures and GANs used in the classification of these structures. The disease diagnosis and image synthesis sections discuss how deep learning models can be used in the early diagnosis of diseases and their applications in medical image synthesis. Finally, the other applications section explores potential new uses of deep learning in fundus image analysis. Additionally, the importance of datasets used for training deep learning models and their impact on model performance is highlighted. This study demonstrates how deep learning techniques can revolutionize retinal image analysis and summarizes technological advancements and methodological innovations in this field. The research also aims to evaluate the impact of these technologies on clinical applications, contributing to the development of diagnostic and treatment methods in the field of eye health.