Background: Diabetic Retinopathy (DR) is an advanced complication of diabetes that damages the retinal blood vessels. If DR is not treated for a long time, it might result in irreversible visual loss or permanent blindness. Problem statement: In the initial stage of DR, the lesions are so minute that the patients do not usually exhibit any symptoms, making early detection challenging. Only specialized eye examinations by experts can monitor the presence of diseases. However, these traditional DR screening methods require a considerable amount of time and labor. Thus, to stop irreversible vision loss, effective and precise automated techniques for early DR identification are required. Objective: This study attempts to capture the recent advancements in DR lesion segmentation, classification and grading using deep learning techniques using fundus images. Methods: Artificial intelligence and computer vision have become effective tools for various automation processes in the health sector including automated disease diagnosis. In the field of DR also, AI- based technologies, especially deep learning has demonstrated promising outcomes by achieving high accuracy for DR detection, stage classification and lesion detection. The paper covers the standardized framework for DR diagnosis using DL, publicly available fundus datasets, preprocessing techniques, DL-based segmentation, classification, and grading methods for DR. Results: Automated DR diagnosis with DL approaches has demonstrated encouraging outcomes. Segmentation approaches such as Mask R-CNN, SegNet, U-Net, and DeepLabv3 have shown high accuracy in DR lesion segmentation. High accuracy rates are also achieved by DL- based classification techniques that employ ensemble models, and transfer learning; with accuracies reported up to 95% and 97% for multi-class and binary classification respectively. Conclusion: Deep learning presents a viable approach to enhance the effectiveness, precision, and accessibility of DR screening. But issues like data availability and model interpretability still exist.