Diabetic Retinopathy (DR) is a prime reason why people lose their sight requiring early detection and accurate classification for DR. In recent times, the development in artificial intelligence and deep learning has significantly enhanced DR detection through automated retinal image analysis systems. This paper examines state-of-the-art techniques which includes models like VGG16, ResNet50, EfficientNetB0, and lightweight Convolution Neural Networks (CNNs), some of which have achieved over 95% accuracy. Techniques like image preprocessing, data augmentation, and Contrast-Limited Adaptive Histogram Equalization (CLAHE) boost performance. New methods such as supervised contrastive learning and Optical Coherence Tomography Angiography (OCTA) based vessel analysis improve diagnosis. However, challenges like poor generalization, unbalanced data, and restricted clinical application still exist. Although, the reviewed models indicate that with the proper deep learning methods and good data handling, highly accurate DR detection is possible.